364
J. Chem. Theory Comput. 2006, 2, 364-382
Design of Density Functionals by Combining the Method of Constraint Satisfaction with Parametrization for Thermochemistry, Thermochemical Kinetics, and Noncovalent Interactions Yan Zhao, Nathan E. Schultz, and Donald G. Truhlar* Department of Chemistry and Supercomputing Institute, UniVersity of Minnesota, 207 Pleasant Street S.E., Minneapolis, Minnesota 55455-0431 Received November 8, 2005
Abstract: We present a new hybrid meta exchange-correlation functional, called M05-2X, for thermochemistry, thermochemical kinetics, and noncovalent interactions. We also provide a full discussion of the new M05 functional, previously presented in a short communication. The M05 functional was parametrized including both metals and nonmetals, whereas M05-2X is a highnonlocality functional with double the amount of nonlocal exchange (2X) that is parametrized only for nonmetals. In particular, M05 was parametrized against 35 data values, and M05-2X is parametrized against 34 data values. Both functionals, along with 28 other functionals, have been comparatively assessed against 234 data values: the MGAE109/3 main-group atomization energy database, the IP13/3 ionization potential database, the EA13/3 electron affinity database, the HTBH38/4 database of barrier height for hydrogen-transfer reactions, five noncovalent databases, two databases involving metal-metal and metal-ligand bond energies, a dipole moment database, a database of four alkyl bond dissociation energies of alkanes and ethers, and three total energies of one-electron systems. We also tested the new functionals and 12 others for eight hydrogen-bonding and stacking interaction energies in nucleobase pairs, and we tested M05 and M05-2X and 19 other functionals for the geometry, dipole moment, and binding energy of HCN-BF3, which has recently been shown to be a very difficult case for density functional theory. We tested eight functionals for four more alkyl bond dissociation energies, and we tested 12 functionals for several additional bond energies with varying amounts of multireference character. On the basis of all the results for 256 data values in 18 databases in the present study, we recommend M05-2X, M05, PW6B95, PWB6K, and MPWB1K for generalpurpose applications in thermochemistry, kinetics, and noncovalent interactions involving nonmetals and we recommend M05 for studies involving both metallic and nonmetallic elements. The M05 functional, essentially uniquely among the functionals with broad applicability to chemistry, also performs well not only for main-group thermochemistry and radical reaction barrier heights but also for transition-metal-transition-metal interactions. The M05-2X functional has the best performance for thermochemical kinetics, noncovalent interactions (especially weak interaction, hydrogen bonding, π‚‚‚π stacking, and interactions energies of nucleobases), and alkyl bond dissociation energies and the best composite results for energetics, excluding metals.
1. Introduction Kohn-Sham density functional theory (DFT) is now one of the most popular tools in the computational and theoretical * Corresponding author phone: (612) 624-7555; fax: (612) 6249390; e-mail:
[email protected].
10.1021/ct0502763 CCC: $33.50
chemistry community, and much progress has been made in the past decade in the development and validation of exchange and correlation functionals.1-67 The line of research developing functionals by requiring them to satisfy constraints has led to the PW91,4 PBE,12 PKZB,23 and TPSS41 functionals on the second and third rungs of “Jacob’s © 2006 American Chemical Society
Published on Web 02/04/2006
Design of Density Functionals
ladder”.30 Although the PKZB functional proved disappointing,23,26,44 the PBE and TPSS functionals have had some notable success in solid-state physics and some areas of chemistry.42,44 However, as pointed out in a prescriptive paper by Perdew et al.,61 PBE and TPSS are not suitable for kinetics (i.e., barrier heights) because both functionals seriously underestimate barrier heights; for example, they were found55 to underestimate barrier height by an average of 8.5 kcal/ mol for 76 barrier heights. The successful DFT methods for kinetics have been developed in a semiempirical way. This involves choosing a flexible functional form depending on one or more parameters and then fitting these parameters to a set of experimental or accurate data. MPW1K,27 BB1K,49 BMK,50 MPWB1K,51 and PWB6K58 are examples of functionals for kinetics determined by the semiempirical approach. The semiempirical approach has also been used to obtain improved functionals for main-group thermochemistry, and a sequence of closely related papers leading successively to functionals called B97,13 B98,16 HCTH,19 B97-1,19 B972,32 τ-HCTH,34 τ-HCTHh,34 BMK,50 and B97-364 provides a good example of this approach. The successive functionals, however, may be improved for one kind of prediction but worsened for another, depending on changes in the functional form, optimization strategy, and training data. A common misconception is that the choice of training data is of overriding importance; actually, the choice of functional form is more critical in that, if the functional form is inadequate, one will not be able to fit a diverse set of data even if it is used for training. Nevertheless, the choice of data is sometimes critical as well. For example, BMK50 is a functional using the same functional form as τ-HCTHh,34 but it was reparametrized against a data set not only for thermochemistry but also for kinetics; the functional form and training set were well enough chosen that BMK performs equally well for kinetics and thermochemistry. However, BMK’s performance for noncovalent interactions is inferior to, for example, PWB6K. PWB6K58 has been shown to be a good functional for weak interactions, and it can describe stacking interactions in small organic clusters59 and nucleobase pairs,60 but its performance for thermochemistry is inferior to that of BMK. It has proved very challenging to develop a functional which can perform well for kinetics, main-group thermochemistry, and noncovalent interactions, including those in nonpolar weakly interacting systems and charge-transfer complexes. It has been stated42 that a “sophisticated nonempirical functional should provide a uniformly accurate description of diverse systems and properties, putting to rest the ‘different functionals for different tasks’ philosophy.” Unfortunately, if one simultaneously considers metallic chemistry and barrier heights in open-shell systems, such a functional did not exist until, in a recent communication,65 we reported a new functional, called M05, which was designed for very general purposes. The M05 functional performs well for all three of the properties mentioned at the end of the previous paragraph and also for transition-metal bond energies, ionization potentials (IPs), and electron affinities (EAs). One purpose of the present paper is to give a more complete account of this new functional. Another purpose is to present
J. Chem. Theory Comput., Vol. 2, No. 2, 2006 365
an alternative parametrization in which transition metals are not included in the training set. The new functional, to be called M05-2X, performs even better for kinetics, thermochemistry, and noncovalent interactions. Since a large number of important applications in chemistry and biochemistry do not involve transition metals, M05-2X may be very useful for such practical work. In contrast, the original M05 functional should be useful for problems involving bonds between two transition metals or metal-ligand bonds where one must treat general metals and organic or inorganic ligands accurately in the same system. In addition, the M05 functional has a fundamental importance in demonstrating the ability of a sufficiently flexible functional form containing kinetic energy density in both the exchange and correlation functionals and parametrized against a purposefully assembled and diverse data set to predict all the data reasonably well. The M05 and M05-2X functionals belong to the fourth rung of Jacob’s ladder (which is explained elsewhere30,41), and they, like the earlier B1B95,10 τ-HCTHh,34 TPSSh,42 BB1K,49 BMK,50 MPW1B95,51 MPWB1K,51 PWB6K,58 PW6B95,58 and TPSS1KCIS functionals,54 can be called hybrid metageneralized gradient approximations (hybrid meta-GGAs), because they incorporate electron spin density, density gradient, kinetic energy density, and Hartree-Fock (HF) exchange. Spin density, density gradient, and kinetic energy density are local properties of the density, although the latter two are sometimes called semilocal (in the early literature, they were sometimes incorrectly called nonlocal), whereas Hartree-Fock exchange is nonlocal. Including Hartree-Fock exchange is sometimes regarded as a temporary expedient that is necessary only because the local exchange-correlation functionals are insufficiently developed, but that is a misimpression. Perdew et al.61 pointed out that, since the exact exchange energy of a fully spin-polarized one-electron system (like a hydrogen atom or H2+) is nonlocal, no local exchange-correlation functional can possibly be correct for this in general (of course, one could force any finite number of one-electron systems to be correct, but this is not the same as getting the effect exactly correct). Thus, the inclusion of Hartree-Fock exchange is a permanent feature of accurate exchange-correlation functionals, not a temporary expedient. The recent post-Hartree-Fock model38,62,63 proposed by Becke employs 100% HartreeFock exchange. One line of argument would be that the ability to tolerate a high percentage of HF exchange and still give good results is the mark of a high-quality density functional. For the present development effort, we combined the semiempirical approach with the incorporation of constraints in the new functionals. The constraints employed are as follows: (1) the new functionals are correct in the uniform electron gas (UEG) limit, and (2) the correlation functional should be free of self-interaction. The first condition is of fundamental importance, and any functional that violates the UEG limit cannot possibly be a universal functional. (Of course, any functional that gets the ionization potential of carbon wrong or the atomization energy (AE) of SiH4 wrong also cannot be universal, but in the present article, the errors
366 J. Chem. Theory Comput., Vol. 2, No. 2, 2006
in ionization potential and atomization energies are minimized with respect to parameter variations, whereas the UEG limit is actually constrained to be exact.) The second constraint is also important even though it does not remove the self-interaction error for the exchange part. Because we use both semiempirical parameter optimization and the method of constraint satisfaction, our approach may be considered to partake of key elements in both of the previously successful lines of functional development. Some workers make a distinction between fitting to analytic results such as fits to the artificial limit of a uniform electron gas or the analytic energy of a hydrogen atom and fitting to numerical results such as the energy of a helium atom, the ionization potential of carbon, or the hydrogen-bond strength of water dimer. Our own philosophy is to use both kinds of information for functional design. Another distinction sometimes made is between using parameters for fitting data and using parameters for shaping a functional. In designing a functional for broad applicability by not only incorporating constraints but also using training data, this distinction becomes arbitrary, and we will not be concerned with it. Section 2 presents our training and test sets. Section 3 gives computational details. Section 4 discusses the theory and parametrization of the new functionals. Section 5 presents results and discusses them. Because the M05 functional was already discussed briefly in a preliminary communication,65 we will discuss the M05-2X functional first.
2. Databases 2.1. M05-2X Training Set. The training set for the M052X models includes the six atomization energies in the AE6 representative database presented previously;68 the binding energies of three dimers,56 (H2O)2, (CH4)2, and (C2H4)2; the binding energy56 of the C2H4‚‚‚F2 charge-transfer complex; the total atomic energies69 of the H, C, O, S, and Si atoms; the ionization potentials66 of C, O, OH, Cu, and Cr; the electron affinities70 of C, O, and OH; the carbon-carbon bond dissociation energies71 of the CH3 bond with CH3 and the isopropyl bond to CH3; and the Kinetics9 database,49,58 which is a database of three forward barrier heights, three reverse barrier heights, and three energies of reaction for the three reactions in the BH668 database. We have previously used Kinetics9 to optimize the BB1K,49 MPWB1K,51 and PWB6K58 methods. Note that we used this small data set to parametrize the new methods, but we assess the new methods with several much larger data sets described below. 2.2. MGAE109/05 Test Set. The MGAE109/05 test set consists of 109 AEs for main-group compounds. All 109 data values are pure electronic energies; that is, zero-point energies and thermal vibrational-rotational energies have been removed by methods discussed previously.54,70,72 The 109 molecules are part of Database/3,72 and the atomization energies of NO, CCH, C2F4, and singlet and triplet CH2 have been updated54 recently. The updated data is a subset of Database/4.73 2.3. Ionization Potential and Electron Affinity Test Set. The zero-point-exclusive IP and EA test sets are called IP13/3 and EA13/3, respectively, and they are taken from a previous
Zhao et al.
paper.70 These data for six atoms and seven molecules are part of Database/3. 2.4. HTBH38/04 Database. The HTBH38/04 database contains 38 transition-state barrier heights for 19 hydrogentransfer (HT) reactions, 18 of which involve radicals as reactants and products. They are taken from previous papers,54,55 and they are also listed in the Supporting Information. 2.5. Noncovalent Interaction Databases. Recently, we developed several databases, in particular, HB6/04,56 CT7/ 04,56 DI6/04,56 WI7/05,58 and PPS5/05,58 for various kinds of noncovalent interactions. HB6/04 is a hydrogen-bond database that consists of the equilibrium binding energies of six hydrogen-bonding dimers, namely, (NH3)2, (HF)2, (H2O)2, NH3‚‚‚H2O, (HCONH2)2, and (HCOOH)2. The CT7/ 04 database consists of the binding energies of seven chargetransfer complexes, in particular, C2H4‚‚‚F2, NH3‚‚‚F2, C2H2‚ ‚‚ClF, HCN‚‚‚ClF, NH3‚‚‚Cl2, H2O‚‚‚ClF, and NH3‚‚‚ClF. The DI6/04 database contains the binding energies of six dipole interaction complexes: (H2S)2, (HCl)2, HCl‚‚‚H2S, CH3Cl‚‚‚HCl, CH3SH‚‚‚HCN, and CH3SH‚‚‚HCl. The WI7/ 05 database consists of the binding energies of seven weak interaction complexes, namely, HeNe, HeAr, Ne2, NeAr, CH4‚‚‚Ne, C6H6‚‚‚Ne, and (CH4)2, all of which are bound by dispersion interactions. The PPS5/05 database consists of binding energies of five π-π stacking complexes, namely, (C2H2)2, (C2H4)2, sandwich (C6H6)2, T-shaped (C6H6)2, and parallel-displaced (C6H6)2. 2.6. Transition-Metal-Transition-Metal and MetalLigand Databases. We employ two databases involving metals. One57 is for the atomization energies of transitionmetal-transition-metal dimers, and it is called the TMAE4/ 05 database; it contains the bond energies of Cr2, Cu2, V2, and Zr2. The other,66 called MLBE4/05, is for the metalligand bond energies in organometallic and inorganometallic complexes, and it contains the Cr-C, Ni-C, Fe-C, and V-S bond energies of CrCH3+, NiCH2+, Fe(CO)5, and VS. These databases are representative subsets of the larger and more diverse TMAE9/0557 and MLBE21/0566 databases. In the present paper, we also use these databases to illustrate the performance of the M05 and M05-2X functionals for the energies of bonds involving metal atoms. 2.7. Alkyl Bond Dissociation Energy (ABDE) Database. This database contains four R-X bond dissociation energies De (R ) Me and X ) CH3 and OCH3). This is called the ABDE4/05 database. The reference D0 values are taken from a recent paper by Izgorodina et al.,71 and we used the B3LYP/ 6-31G(d) zero-point vibrational energies scaled with a scale factor of 0.980674 to obtain De. 2.8. Dipole Moment Database. This database consists of the fixed-geometry dipole moments for six molecules, namely, N6, H2CO, CuH, BF, LiCl, and H2O, where N6 is R-amino,ω-nitro-dodecahexaene, which has the formula H2N(CHdCH)6NO2. This database is called the DM6/05 database. We use the MP2/6-31G geometry75 for the N6 molecule, and the reference dipole moment is computed at the MP2/ 6-311+G(2df,2p) level of theory since previous work76 showed good agreement between the MP2 and CCSD(T) levels of theory for a smaller basis set. For the
Design of Density Functionals
CuH molecule, we use the geometry from the modified coupled pair functional (MCPF) calculations of Langhoff and Bauschlicher.77 The reference dipole moment for CuH is an average of the values (2.95 and 2.98 D, respectively) obtained by their MCPF calculation77 and our own65 CCSD(T)/ANO calculation, where ANO denotes the triple-ζ atomic natural orbital basis set of Widmark et al.78,79 The geometries and accurate dipole moments for H2CO, BF, LiCl, and H2O are calculated at the CCSD(T)/aug-cc-pVTZ level of theory. 2.9. IPEA8 Database. The IPEA8 database contains the ionization potentials of C, O, OH, Cr, and Cu and the electron affinities of C, O, and OH. 2.10. AAE5 and AAE4 Databases. The AAE5 database consists of the total atomic energies69 of H, C, O, S, and Si, and the AAE4 database is the same as AAE5 except that it excludes the atomic energy of H.
3. Computational Methods 3.1. Geometries, Basis Sets, and Spin-Orbit Energy. All calculations for the AE6, MGAE109/05, IP13/3, EA13/3, and HTBH38/04 databases are single-point calculations at QCISD/ MG3 geometries, where QCISD is quadratic configuration interaction with single and double excitations80 and MG3 is the modified81,82 G3Large83 basis set. The MG3 basis set,81 also called G3LargeMP2,82 is the same as 6-311++G(3d2f,2df,2p)84,85 for H-Si but improved83 for P-Ar. The geometries for all of the molecules in the HB6/04, CT7/04, DI6/04, and WI7/05 noncovalent databases and the (C2H4)2 and (C2H2)2 dimers in the PPS5/05 database are optimized at the MC-QCISD/3 level, where MC-QCISD is the multicoefficient QCISD method.72,86 The geometries for the benzene dimers in the PPS5/05 database are taken from Sinnokrot and Sherrill.87 The geometries for all of the molecules in the ABDE4/05 database are optimized at the B3LYP/6-31G(d) level, and they are taken from the Supporting Information of a previous paper.71 The 6-311+G(3df,2p) basis set is used for the calculations of ABDEs for the purpose of comparison with the previous results. The geometries for the molecules in the transition-metaltransition-metal (TMAE4/05) and metal-ligand (MLBE4/ 05) databases are optimized consistently with each level of theory. We used the double-ζ-quality DZQ basis set57 for the calculations on the molecules in these two databases. The DZQ basis set uses the relativistic effective core potential method of Stevens et al.88 for both the 3d and 4d transition metals, and it uses the 6-31+G(d,p) basis set for main-group atoms. In these cases (i.e., for the metal-compound calculations), the d functions are spherical harmonic 5D sets. Although one requires triple-ζ quality or better basis sets for quantitative results on transition metals, DZQ is good enough for a broad survey of many functionals to ascertain which ones gives relatively good results for bonds involving metal atoms. The geometries for the stacked and hydrogen-bonded nucleobase pairs are optimized at the PWB6K/6-31+G(d,p) level. All DFT calculations for the base pairs use the 6-31+G(d,p) basis set.
J. Chem. Theory Comput., Vol. 2, No. 2, 2006 367
To test the functionals for the one-electron systems, we employed the cc-pVQZ basis set for the hydrogen atom, H2+ (re ) 1.4 bohr), and H2+ (re ) 2.0 bohr). For the DM6/05 dipole moment database, we used the TZQ basis set, which is described in our previous paper.57,66 For the MGAE109, HTBH38/04, IP13/3, EA13/3, and all five noncovalent databases, we used the MG3S basis sets for single-point energy calculations. The MG3S basis70 is the same as MG3 except it omits diffuse functions on hydrogens. Note that all of the basis sets mentioned above use pure d or f functions except the 6-31+G(d,p) basis set employed in the calculations for nucleobase pairs, which uses Cartesian basis functions. In all of the calculations presented in this paper, the spinorbit stabilization energy was added to atoms and open-shell molecules for which it is nonzero, as described previously.57,66,81 3.2. Counterpoise Correction. For noncovalent complexes, we perform calculations with and without the counterpoise corrections89,90 for the basis set superposition error (BSSE). 3.3. Software. All of the calculations were performed with a locally modified version of the Gaussian03 program91 except that the benchmark CCSD(T) calculations of the dipole moment for CuH were calculated with MOLPRO.92
4. Theory and Parametrization 4.1. Meta-GGA Exchange Functional. The functional form adopted for the meta-GGA exchange functional is E(0) X )
PBE (Fσ,∇Fσ) f(wσ) ∑σ ∫ dr F Xσ
(1)
PBE (Fσ,∇Fσ) is the exchange energy density of the where F Xσ PBE11 exchange model (which has the same functional form as the earlier exchange functional of Becke,1 but with different values for the two parameters) and f(wσ) is the kinetic-energy-density enhancement factor m
f(wσ) )
aiwiσ ∑ i)0
(2)
where the variable wσ is a function of tσ, and tσ is a function of the kinetic energy density τσ of electrons with spin σ. wσ ) (tσ - 1)/(tσ + 1)
(3)
tσ ) τLSDA /τσ σ
(4)
where
τσ )
1occup 2
≡ τLSDA σ
∑i |∇Ψiσ|2
3 (6π2)2/3Fσ5/3 10
(5)
(6)
The motivation for the functional form in eqs 1-6 is explained in our previous paper,65 and here we simply emphasize the key elements, namely, that it allows us to combine the correct UEG limit with reasonable behavior for
368 J. Chem. Theory Comput., Vol. 2, No. 2, 2006
Zhao et al.
a large reduced density gradient and with Becke’s strategy18,25 for simulating delocalized exchange by local density functionals by using local functionals to detect delocalization and inhomogeneity. 4.2. Meta-GGA Correlation Functional. In the correlation functional, we treat the opposite-spin and parallel-spin correlations differently. We begin with Perdew and Wang’s functional5 for the correlation part of the local spin density approximation (LSDA). Then, following the analysis of Stoll et al.,93 one can decompose the LSDA correlation energy into opposite-spin (denoted Rβ) and parallel-spin (denoted σσ, RR, and ββ, depending on the content) correlation energy components for the UEG:
proposed by Becke18), which gives the right UEG limit but does not have the above-mentioned numerical instability. The opposite-spins correlation energy of our new functional is expressed as ERβ C )
(8)
(FR,Fβ) is the LSDA correlation energy. Rewhere ELSDA C cently, Gori-Giorgi et al.94 showed that the spin resolution of the uniform electron gas correlation energy by eqs 7 and 8 is not accurate for spin-unpolarized (FR ) Fβ) systems. More recently, Gori-Giorgi and Perdew proposed a better formula.95 Note that eq 8 does not vanish in the one-electron case, and this nonvanishing is a manifestation of self-interaction error. To correct this self-interaction error, Becke10 used a quantity, Dσ, which is defined as Dσ ) 2τσ -
1 |∇Fσ| 4 Fσ
2
(9)
where τσ is the kinetic energy density of electrons with spin σ, defined in eq 10. The function Dσ can also be written as Dσ ) 2(τσ - τW σ)
(10)
¨ cker kinetic energy density96 where τW σ is the von Weizsa given by τW σ )
2 1 |∇Fσ| 8 Fσ
(11)
In a one-electron case, τσ ) τW σ , so Dσ vanishes in any oneelectron system. Note that the uniform electron gas limit (∇Fσf 0) of Dσ is 3 ) (6π2)2/3F5/3 DUEG σ σ 5
(12)
as a self-interaction correction factor Becke used Dσ/DUEG σ to the parallel-spin case for the B95 correlation functional.10 We have pointed out previously56 that the function DUEG in σ the denominator causes some self-consistent field convergence problems that can be eliminated by using a different cutoff criterion. The DUEG in the denominator also causes σ some integration grid problems as pointed by Johnson and co-workers.53 To avoid these numerical problems, we used a different self-interaction correction factor, Dσ/2τσ (also
(13)
where gRβ(xR,xβ) is defined as n
gRβ(xR,xβ) )
cCRβ,i ∑ i)0
[
γCRβ(x2R + x2β)
1 + γCRβ(x2R + x2β)
]
i
(14a)
where
UEG ECRβ (FR,Fβ) ) ELSDA (FR,Fβ) - ELSDA (FR,0) - ELSDA (0,Fβ) C C C (7) UEG ECσσ (FR) ) ELSDA (FR,0) C
∫eRβUEG gRβ(xR,xβ) dr
xσ )
|∇Fσ|
σ ) R, β
F4/3 σ
(14b)
For parallel spins, Eσσ C )
D
σ dr ∫eUEG σσ gσσ(xσ) 2τσ
(15)
where Dσ/2τσ is the self-interaction correction factor and n
gσσ(xσ) )
cCσσ,i ∑ i)0
(
γCσσx2R
)
i
1 + γCσσx2R
(16)
UEG Note that eRβ and eUEG σσ in eq 13 and eq 15 are the UEG correlation energy density for the antiparallel-spin and parallel-spin cases, respectively, and they can be extracted from the total UEG correlation energy density in the same way as shown in eqs 7 and 8. The total correlation energy of the new correlation functional is given by RR ββ EC ) ERβ C + E C + EC
(17)
Note that our new correlation functional is similar to the correlation functional in the BMK50 method; the difference is that BMK does not have the self-interaction correction factor Dσ/2τσ for the parallel-spin case. We require cCRβ,0 ) cCσσ,0 ) 1 in eqs 14a and 16. In agreement with the philosophy of the B95 functional,10 this forces the correlation functionals to have the correct UEG limit, which is not enforced in a considerable body of work16,19,34,50 using similar correlation functionals. One can easily confirm that our new correlation functional gives the right UEG limit (with xσ f 0, Dσ f 2τσ f DUEG σ ). Following Becke,18 we preoptimized the γ parameters to the correlation energies of He and Ne in a preliminary fit. The values of these two nonlinear parameters in the new functionals are γCRβ ) 0.0031 and γCσσ ) 0.06
(18)
4.3. Hybrid Meta Functional. The hybrid exchangecorrelation energy can be written as follows: Ehyb XC )
X HF X DFT E + EDFT E + 1C 100 X 100 X
(
)
(19)
where EHF X is the nonlocal Hartree-Fock exchange energy,
Design of Density Functionals
J. Chem. Theory Comput., Vol. 2, No. 2, 2006 369
Table 1. Optimized Parameters in the M05-2X and M05 Methods M05-2X
M05
parameters
ai
cCRβ,i
cCσσ,i
ai
cCRβ,i
cCσσ,i
0 1 2 3 4 5 6 7 8 9 10 11 X
1.000 00 -0.568 33 -1.300 57 5.500 70 9.064 02 -32.210 75 -23.732 98 70.229 96 29.886 14 -60.257 78 -13.222 05 15.236 94
1.000 00 1.092 97 -3.791 71 2.828 10 -10.589 09
1.000 00 -3.054 30 7.618 54 1.476 65 -11.923 65
1.000 00 0.081 51 -0.439 56 -3.224 22 2.018 19 8.794 31 -0.002 95 9.820 29 -4.823 51 -48.175 74 3.648 02 34.022 48
1.000 00 3.785 69 -14.152 61 -7.465 89 17.944 91
1.000 00 3.773 44 -26.044 63 30.699 13 -9.226 95
56
28
X is the percentage of Hartree-Fock exchange in the hybrid functional, EDFT is the local DFT exchange energy, and X EDFT is the local DFT correlation energy. Equation 19 can C be rewritten as
(
HF Ehyb XC ) EX + 1 -
X DFT (EDFT - EHF X ) + EC 100 X
)
(20)
From eq 20, one can see that the total correlation energy for a DFT calculation is modeled as the sum of the dynamic correlation energy given by EDFT and the nondynamical C - EHF correlation energy29 contained in (1 - X/100)(EDFT X X ). We optimize X along with the parameters in the new meta exchange and correlation functionals; the optimization procedure is given in the next section. 4.4. Optimization of the New Hybrid Meta-GGA. All of the parameter optimizations were carried out with a genetic algorithm.97 The parameters ai in eq 2 are determined by fitting them to the data in the training set with a constraint that a0 ) 1, which enforces the UEG limit. This limit corresponds to tσ = 1, wσ = 0, and xσ = 0; and f(wσ) should tend to unity in this limit because the PBE exchange functional satisfies the UEG limit. Therefore, we constrained a0 to unity to enforce this limit. Simultaneously, we optimized the cCRβ,i and cCσσ,i parameters in eqs 14a and 16 to the data in the training set. The M05 and M05-2X functionals were optimized using different training sets. In both new methods, we optimize the ai parameters in the exchange functional, the cCRβ,i and cCσσ,i parameters in the correlation functional, and the percentage, X, of Hartree-Fock exchange. We minimize the training function with respect to these parameters in a selfconsistent way by solving the Fock-Kohn-Sham equation using the basis set and geometries described in section 3.1. We optimized the parameters in M05-2X against the data in the training set to minimize the following training function F ) RMSEPB(AE6) + RMSE(IPEA8) + RMSE(Kinetics9) + 10 × RMSE(NB4) + RMSE(ABDE2) + 0.2 × RMSE(AAE4) + 2 × UE(AEH) (21) where RMSEPB is the root-mean-squared error (RMSE) per bond. In particular, RMSEPB is obtained by dividing the RMSE for the AE6 database by the average number of bonds per molecule in this database. The second term is the RMSE
for the IPEA8 database, which is defined in section 2. The third term is the RMSE for the Kinetics9 database. RMSE(NCCE4) is the RMSE for four noncovalent complexation energies, namely, the equilibrium binding energies of the (H2O)2, (CH4)2, and (C2H4)2 dimers and that of the C2H4‚‚ ‚F2 charge-transfer complex; RMSE(ABDE2) is the rootmean-square error in the bond dissociation energies of CH3CH3 and isopropyl-CH3; RMSE(AAE4) is the RMSE for the total atomic energies of C, O, S, and Si; and UE(AEH) is the unsigned error in the atomic energy of the hydrogen atom. Our preliminary fitting showed that, for the training function F, nonphysical parameters are produced when m > 11 or n > 4, so we used m ) 11 in eq 2 and we used n ) 4 in eqs 14a and 16. Thus, we optimized 20 parameters (11 in ai, 4 in cCRβ,i, and 4 in cCσσ,i and X) for the M05-2X method. All optimized parameters for M05-2X are listed in Table 1 along with the parameters for the M05 functional. In the optimization of the M05 functional,65 the RMSE(ABDE2) is replaced by the RMSE for the bond dissociation energies of Cr2 and V2 and the Cr-C bond of CrCH3+, and the weight we used for the error for the atomic energy of hydrogen is 0.2 instead of 2. In the original work on the M05 functional, we found that we could obtain very similar results (the mean unsigned error for nonmetals was about 1% smaller and that for metals was about 13% larger) by employing the same strategy with the PBE exchange functional replaced by the mPW15 one. Thus, the treatment of kinetic energy density and correlation energy, along with the consistency between the exchange and correlation functionals, is the key ingredients in the M05 functionals, not the precise form of Fxσ. The optimized functions of eq 2 for the final M05 and M05-2X functionals are shown in Figure 1. A useful way to visualize the meta-GGA nonlocality is to write the meta-GGA exchange-correlation energy as EXC[FR,Fβ] )
∫ d3r FUEG X (F) FXC(FR,Fβ,∇FR,∇Fβ,τR,τβ) (22)
) where F ) FR + Fβ is the total density and -(3/4π)(3π2F) is the exchange energy per electron of a spinunpolarized (FR ) Fβ) uniform electron gas; the enhancement factor FXC shows the effects of correlation and inhomogeneity.45 To visualize FXC for the meta-GGA part of the M05 and M05-2X functionals, we define three quantities, namely, UEG X
370 J. Chem. Theory Comput., Vol. 2, No. 2, 2006
Zhao et al.
Figure 1. The τ enhancement factors for the M05 and M052X functionals.
Figure 2. M05 enhancement factor FXC of eq 22 as a function of the reduced gradient s of eq 23 with Rτ ) 0.2 for various spin-unpolarized (FR ) Fβ) densities ranging from the highdensity (rs ) 0) to the exchange-only limit (rs f ∞).
Figure 3. M05 enhancement factor FXC of eq 22 as a function of the reduced gradient s of eq 23 with Rτ ) 1 for various spin-unpolarized (FR ) Fβ) densities ranging from the highdensity (rs ) 0) to the exchange-only limit (rs f ∞).
Figure 4. M05-2X enhancement factor FXC of eq 22 as a function of the reduced gradient s of eq 23 with Rτ ) 0.2 for various spin-unpolarized (FR ) Fβ) densities ranging from the high-density (rs ) 0) to the exchange-only limit (rs f ∞).
FXC(rs,s) e 2.273
s, rs, and Rτ |∇F| s) (24π2)1/3F4/3 rs )
Rτ )
(4πF3 )
(23)
1/3
(24)
τσ - τW σ
(25)
τLSDA σ
We note, though, that these figures only show the behavior of the meta-GGA part of the functional, and we do not recommend users to use the pure meta-GGA part of the M05 or M05-2X method, because the parameters of both functionals are optimized with the mixing of a certain amount of the Hartree-Fock exchange. We are working on the optimization of a pure local meta-GGA without the HartreeFock exchange using the same functional form.
5. Results and Discussion
By using eqs 6, 11, 23-25, we can transform the kinetic energy density into a function of s, rs, and Rτ. Figures 2-5 show the enhancement factors for the meta-GGA part of the M05 and M05-2X functionals. Figures 2-5 show that both functionals violate the scaling inequality:98 FXC(r′s,s) > FXC(rs,s)
(27)
r′s > rs
and they also violate the Lieb-Oxford bound99
(26)
5.1. Assessment of the New Hybrid Meta Functionals. We fitted our new functionals against a small and diverse data set (six data values in AE6, nine data values in Kinetics9, four data values for noncovalent complexation, eight data values for ionization energies and electron affinities, and three data values for transition-metal-transition-metal and metal-ligand interactions), but we assess the new functionals against a much larger data set that includes 109 main-group atomization energies, 13 ionization potentials (IP13), 13
Design of Density Functionals
J. Chem. Theory Comput., Vol. 2, No. 2, 2006 371
for each of the individual databases and the five smallest average errors overall (for each table) are in bold. 5.2. Thermochemistry: AE, IP, and EA Results. Table 3 summarizes the errors in AEs, IPs, and EAs for all of the tested methods. Table 3 shows that the PW6B95, M05-2X, and BMK methods give the best results for AE calculations, and they give a MUEPB of less than 0.5 kcal/mol. MPWB1K, BB1K, and MPW1B95 have the best performance for IP calculations, whereas BMK, PW6B95, and τ-HCTHh give the best performance for EA calculations. To compare their performance for thermochemistry, we defined the TMUE (total MUE) as the mean signed error over all 135 data values in this table:
Figure 5. M05-2X enhancement factor FXC of eq 22 as a function of the reduced gradient s of eq 23 with Rτ ) 1 for various spin-unpolarized (FR ) Fβ) densities ranging from the high-density (rs ) 0) to the exchange-only limit (rs f ∞)
electron affinities (EA13), 38 barrier heights (HTBH38), 19 energies of reaction, 6 hydrogen-bonding energies (HB6), 7 charge-transfer complexation energies (CT7), 6 complexation energies of complexes dominated by dipole interaction (DI6), 7 weak interaction energies (WI7), 5 π-π stacking interaction energies (PPS5), 4 transition-metal-transition-metal bond energies (TMAE4), 4 metal-ligand bond energies (MLBE4), 6 dipole moments, 4 alkyl bond dissociation energies (R-CH3 and R-OCH3), and 8 nucleobase pair interaction energies. We compare the results obtained by the new methods to those for 28 other functionals. Table 2 lists all 30 density functionals considered in this work. In each case, we specify the year it was first published, the functional forms used for the dependence on ∇F, whether the functional includes τ in the exchange and correlation functional, and whether the correlation functional is self-correlation-free (SCorF). Table 2 also contains two columns (one for the exchange functional and one for the correlation functional) that tell whether the functional reduces to the correct uniform electron gas limit when ∇Fσ f 0 and τσ f τLSDA . σ In most of the comparisons, we will gauge the quality of the results by mean unsigned errors (MUEs), which are the averages of the absolute deviations of the calculated values from database reference values, and by mean signed errors (MSEs), which are used to detect systematic deviations. However, for atomization energies, we use MUE per bond (MUEPB) and MSE per bond (MSEPB) because this allows46,49 a more transferable comparison between databases with different average sizes of molecules. Because the dipole moments considered in the dipole moment database vary widely in magnitude, we also consider mean signed percentage error (MS%E) and mean unsigned percentage error (MU%E). We also use MU%E in one later table because the quantities considered in that table do not all have units of energy. To make the trends more clear, in every table, we will list the methods in increasing order of the values in the key (overall) error column, which is always the last column of a given table. The five smallest average errors
TMUE ) [MUEPB(AE) × 109 + MUE(IP) × 13 + MUE(EA) × 13]/135 (28) If we use TMUE as a criterion of practical usefulness for thermochemistry, Table 3 shows that M05-2X is the best functional, followed by PW6B95 and BMK. 5.3. Thermochemical Kinetics. Table 4 gives the mean errors for the HTBH38/04 database. A total of 18 of the 19 reactions in this database involve radicals as reactants or products, and 16 of those involve an odd number of electrons. Systems with an odd number of electrons and stretched bonds are well-known to provide a critical test case for density functional theory. Furthermore, we have shown elsewhere55 that functionals that perform well for hydrogen-transfer barrier heights also perform well for barrier heights of more general classes of reaction, so we believe that good performance on this database is critical if a functional is to be judged as broadly applicable. Table 4 shows that BB1K gives the best results for barrier heights, with PWB6K, MPWB1K, MPW1K, BMK, and M05-2X being less accurate on average by 0.12-0.18 kcal/mol. M05 and the very new B97-3 are less accurate than these six functionals by 0.59-0.77 and 0.93-1.11 kcal/mol, respectively; whereas, the other 22 functionals in the table are less accurate than these six by 1.46-15.56 kcal/mol. The mean unsigned error in the energies of reaction for the 19 reactions is called MUE(∆E19) and is given in the second-to-last column of Table 4. M05-2X gives the best performance for these energies of reaction, followed by B1B95, PW6B95, MPW1B95, BMK, and M05. Right behind these six are B98, τ-HCTHh, B3PW91, B97-2, and mPW1PW91, with the other 19 functionals being significantly less accurate. We also tabulated an average MUE (called AMUE) that is defined as AMUE ) [MUE(∆E19) + MMUE(BH38)]/2
(29)
where MUE(∆E19) is the mean unsigned error for the energy of reactions for the 19 reactions in the HTBH38 database. If we use AMUE as a criterion to judge the performance of a DFT method for thermochemical kinetics, Table 4 shows that M05-2X, BMK, BB1K, MPW1K, MPWB1K, and M05 are the best methods for kinetics. Table 4 is particularly encouraging in that M05-2X has a mean unsigned error for hydrogen-transfer barrier height on the order of 1.3 kcal/ mol, a level of accuracy that is significantly exceeded only by the BB1K density functional, which is much less broadly applicable, and by large-basis CCSD(T) or some equally
372 J. Chem. Theory Comput., Vol. 2, No. 2, 2006
Zhao et al.
Table 2. Density Functionals exchange
correlation
method
year
ref(s)
∇F
X
τ?
UEG?
∇F
τ?
SCorF?
UEG?
BLYP SPWL B3PW91 B3LYP BB95 B1B95 G96LYP PBE B1LYP mPWPW91 mPW1PW91a B98 B97-1 PBE1PBEb MPW1K B97-2 τ-HCTHh TPSS TPSSh X3LYP BB1K BMK MPW1B95 MPWB1K TPSS1KCIS PW6B95 PWB6K B97-3 M05 M05-2X
1988 1992 1993 1994 1996 1996 1996 1996 1997 1998 1998 1998 1998 1999 2000 2001 2002 2003 2003 2004 2004 2004 2004 2004 2005 2005 2005 2005 2005 2005
2, 3 5, 144 2, 4, 7 2, 3, 8 2, 10 2, 10 3, 9 11 2, 3, 14 4, 15 4, 15 16 19 22 27 19 34 41 42 3, 47 2, 10, 49 50 10, 15, 51 10, 15, 51 21, 41, 60 58 58 64 65 present
B88 Slater B88 B88 B88 B88 G96 PBE B88 mPW mPW B98 B97-1 PBE mPW B97-2 τ-HCTHh TPSS TPSS X B88 BMK mPW mPW TPSS PW6B95 PWB6K B97-3 M05 M05-2X
0 0 20 20 0 28 0 0 25 0 25 21.98 21 25 42.8 21 15 0 10 21.8 42 42 31 44 13 28 46 26.93 28 56
no no no no no no no no no no no no no no no no yes yes yes no no yes no no yes no no no yes yes
yes yes yes yes yes yes yes yes yes yes yes no no yes yes no no yes yes yes yes no yes yes yes yes yes no yes yes
LYP PW91-L PW91 LYP B95 B95 LYP PBE LYP PW91 PW91 B98 B97-1 PBE PW91 B97-2 τ-HCTHh TPSS TPSS LYP B95 BMK B95 B95 KCIS PW6B95 PWB6K B97-3 M05 M05-2X
no no no no yes yes no no no no no no no no no no no yes yes no yes no yes yes yes yes yes no yes yes
yes no no yes yes yes yes no yes no no no no no no no no yes yes yes yes no yes yes yes yes yes no yes yes
no yes yes no yes yes yes yes no yes yes no no yes yes no no yes yes no yes no yes yes yes yes yes no yes yes
a
Also called mPW0 and MPW25. b Also called PBE0.
Table 3. Mean Errorsa [kcal/mol for Ionization Potentials (IP) and Electron Affinities (EA) and kcal/mol per Bond for Atomization Energies (AE)] MGAE109/05 method PW6B95 M05-2X BMK B1B95 MPW1B95 M05 B98 B97-3 B97-2 TPSS1KCIS B97-1 B3PW91 τ-HCTHh PBE1PBE mPW1PW91 TPSS TPSSh MPWB1K B3LYP BB1K PWB6K X3LYP BLYP mPWPW91 G96LYP BB95 MPW1K B1LYP PBE SPWL averageb
IP13/3
EA13/3
MSEPB MUEPB MSE MUE MSE MUE TMUE -0.02 -0.02 -0.04 -0.23 0.31 -0.01 -0.50 -0.37 -0.20 -0.05 -0.39 -0.13 -0.21 0.11 -0.73 0.63 -0.12 -0.84 -0.69 -1.32 -1.41 -1.26 -0.47 1.72 -1.39 2.18 -2.33 -2.66 2.80 16.89
0.40 0.48 0.47 0.55 0.62 0.53 0.64 0.59 0.65 0.67 0.75 0.66 0.75 0.91 0.88 1.03 0.98 0.98 0.91 1.34 1.43 1.42 1.49 2.01 1.96 2.34 2.34 2.69 3.03 16.89 1.68
2.24 1.69 2.74 -0.13 0.36 -0.41 1.99 1.56 0.46 0.91 0.99 3.70 3.62 2.44 3.17 1.80 1.96 0.51 3.58 0.13 1.57 2.58 -0.41 2.93 -1.12 -0.55 3.41 -0.13 2.11 4.34
3.24 3.54 4.21 2.18 2.14 2.87 3.21 3.51 2.21 2.63 2.84 4.25 4.03 3.23 3.72 3.11 3.17 2.05 4.72 2.09 2.28 4.73 4.87 4.15 4.64 3.34 3.53 3.80 3.58 5.18 3.43
0.72 0.53 0.28 3.02 2.72 2.81 0.30 0.82 2.41 1.84 1.09 -0.12 -1.18 1.50 1.09 0.51 1.40 3.99 -1.51 4.28 3.23 -0.41 -0.11 -1.56 1.33 0.21 2.79 2.56 -1.20 -5.77
1.78 2.03 1.56 3.16 2.91 2.96 1.84 2.07 2.89 2.81 2.02 2.09 1.83 2.76 2.62 2.31 2.81 4.11 2.29 4.36 3.59 3.04 2.63 2.26 2.93 1.99 3.71 3.64 2.22 5.80 2.77
0.81 0.93 0.94 0.96 0.98 0.99 1.00 1.02 1.02 1.07 1.07 1.14 1.17 1.31 1.32 1.36 1.37 1.38 1.41 1.70 1.72 1.89 1.93 2.24 2.31 2.40 2.59 2.89 3.01 14.70 1.95
a
MUEPB denotes mean unsigned error (MUE) per bond. MSE denotes mean signed error. TMUE denotes total MUE, and it is defined as TMUE ) [MUEPB × 109 + MUE(IP) × 13 + MUE(AE) × 13]/135. b In all tables, where the last row is “average”, it is the average of that column for all functionals in the table.
expensive wave function theory. Furthermore, both M052X and M05 are in the top seven for each of these three
Table 4. Mean Errors for Thermochemical Kineticsa,b HTBH38/04
∆E19
methods
X
MSE
MUE
MUE
AMUEc
M05-2X BMK BB1K MPW1K MPWB1K PWB6K M05 B97-3 B1B95 MPW1B95 PW6B95 B97-2 mPW1PW91 B3PW91 B98 B1LYP PBE1PBE B97-1 B3LYP τ-HCTHh TPSS1KCIS X3LYP G96LYP TPSSh BB95 BLYP TPSS mPWPW91 PBE SPWL Average
56 42 42 42.8 44 46 28 26.93 28 31 28 21 25 20 21.98 25 25 21 20 15 13 21.8 0 10 0 0 0 0 0 0
-0.39 -0.82 -0.57 -0.60 -0.85 -0.50 -1.20 -2.11 -2.80 -3.02 -3.14 -3.09 -3.54 -4.02 -4.16 -2.84 -4.22 -4.40 -4.13 -5.29 -4.69 -3.98 -6.25 -5.97 -8.14 -7.52 -7.71 -8.43 -9.32 -17.72
1.34 1.32 1.16 1.32 1.29 1.28 1.93 2.27 2.80 3.02 3.14 3.24 3.55 4.03 4.16 3.18 4.22 4.40 4.23 5.29 4.69 4.09 6.26 5.97 8.14 7.52 7.71 8.43 9.32 17.72 4.57
0.64 0.92 1.38 1.31 1.41 1.57 0.95 1.15 0.78 0.86 0.85 1.08 1.13 1.05 0.97 2.29 1.29 1.48 1.95 0.97 1.64 3.03 2.26 2.65 1.63 2.29 2.53 1.97 2.71 6.39 1.70
0.99 1.12 1.27 1.32 1.35 1.42 1.44 1.71 1.79 1.94 1.99 2.16 2.34 2.54 2.57 2.73 2.76 2.94 3.09 3.13 3.16 3.56 4.26 4.31 4.89 4.90 5.12 5.20 6.01 12.05 3.14
a The MG3S basis used for all calculations in this table. b MUE denotes mean unsigned error (kcal/mol). MSE denotes mean signed error (kcal/mol). c AMUE in this table is calculated by averaging the two MUE columns, and it is a measure of the quality of a method for kinetics.
mean unsigned error columns in Table 4. The only other functional that appears in the top-seven list for all three of these columns is BMK.
Design of Density Functionals
J. Chem. Theory Comput., Vol. 2, No. 2, 2006 373
Table 5. Mean Errors for Noncovalent Databases (kcal/mol)a,b,c HB6/04 MUE
CT7/04
MMUE
method
no-cp
cp
M05-2X PWB6K M05 MPWB1K PW6B95 MPW1B95 B97-1 PBE1PBE B98 MPW1K X3LYP mPW1PW91 TPSS1KCIS TPSSh BMK B3LYP BB1K PBE B1LYP B97-3 TPSS B97-2 B1B95 mPWPW91 B3PW91 τ-HCTHh BLYP BB95 SPWL G96LYP average
0.40 0.44 0.58 0.41 0.53 0.50 0.45 0.40 0.45 0.33 0.45 0.39 0.49 0.41 0.68 0.60 0.99 0.45 0.72 1.16 0.45 1.22 1.31 0.57 1.03 1.94 1.18 1.83 3.13 2.95 0.88
0.20 0.34 0.53 0.70 0.78 0.86 0.45 0.28 0.66 0.61 0.48 0.79 0.86 0.80 0.96 0.93 1.37 0.32 1.05 1.50 0.82 1.64 1.69 0.96 1.43 2.58 1.56 2.21 2.67 3.30 1.11
0.30 0.39 0.55 0.56 0.65 0.68 0.45 0.34 0.55 0.47 0.47 0.59 0.67 0.60 0.82 0.76 1.18 0.39 0.88 1.33 0.63 1.43 1.50 0.77 1.23 2.26 1.37 2.02 2.90 3.13 1.00
MUE
DI6/04 MMUE
no-cp
cp
0.46 0.25 0.68 0.24 0.69 0.47 1.17 1.04 0.91 0.44 0.96 0.65 1.22 1.44 0.41 0.71 0.68 2.95 0.49 0.48 2.20 0.56 0.53 2.25 0.64 1.60 1.67 1.48 5.61 1.20 1.14
0.30 0.16 0.30 0.45 0.47 0.31 0.89 0.75 0.66 0.66 0.68 0.51 0.95 1.16 0.62 0.54 1.00 2.63 0.45 0.63 1.86 0.67 0.72 1.89 0.69 1.42 1.42 1.27 5.23 1.28 1.02
0.38 0.21 0.49 0.34 0.58 0.39 1.03 0.90 0.79 0.55 0.82 0.58 1.08 1.30 0.52 0.63 0.84 2.79 0.47 0.56 2.03 0.61 0.62 2.07 0.66 1.51 1.54 1.38 5.42 1.24 1.08
MUE
WI7/05 MMUE
no-cp
cp
0.27 0.24 0.23 0.50 0.40 0.50 0.28 0.35 0.34 0.52 0.45 0.53 0.46 0.49 0.78 0.78 1.02 0.46 0.93 0.82 0.52 0.87 1.11 0.56 0.97 0.75 1.00 1.18 2.16 2.57 0.73
0.32 0.32 0.24 0.65 0.49 0.63 0.30 0.38 0.40 0.67 0.59 0.63 0.55 0.58 0.97 0.94 1.16 0.40 1.09 0.98 0.56 1.02 1.26 0.59 1.14 1.01 1.18 1.35 1.95 2.74 0.84
0.29 0.28 0.23 0.57 0.45 0.56 0.29 0.37 0.37 0.60 0.52 0.58 0.50 0.54 0.88 0.86 1.09 0.43 1.01 0.90 0.54 0.94 1.19 0.57 1.06 0.88 1.09 1.27 2.05 2.65 0.79
MUE
PPS5/05 MMUE
no-cp
cp
0.09 0.15 0.14 0.08 0.11 0.10 0.10 0.12 0.12 0.20 0.16 0.58 0.17 0.18 0.76 0.31 0.34 0.13 0.30 0.49 0.19 0.25 0.42 0.24 0.53 0.44 0.45 0.57 0.20 1.37 0.31
0.03 0.07 0.06 0.16 0.09 0.16 0.11 0.18 0.16 0.29 0.22 0.30 0.21 0.26 0.85 0.39 0.44 0.15 0.39 0.58 0.26 0.35 0.51 0.32 0.62 0.33 0.53 0.66 0.10 1.47 0.34
0.06 0.11 0.10 0.12 0.10 0.13 0.10 0.15 0.14 0.25 0.19 0.44 0.19 0.22 0.81 0.35 0.39 0.14 0.35 0.53 0.22 0.30 0.47 0.28 0.58 0.38 0.49 0.62 0.15 1.42 0.33
MUE
MMUE MMMUE
no-cp
cp
0.49 0.79 1.12 1.32 1.21 1.46 1.57 1.84 1.91 2.25 2.49 2.43 2.39 2.46 2.36 2.95 2.03 1.86 3.06 2.49 2.53 2.73 2.34 2.69 3.23 2.11 3.58 2.96 0.35 5.19 2.21
0.71 1.00 1.34 1.57 1.44 1.70 1.78 2.09 2.13 2.53 2.71 2.71 2.62 2.72 2.57 3.17 2.27 2.09 3.27 2.70 2.78 2.96 2.58 2.96 3.49 2.37 3.79 3.18 0.43 5.41 2.43
0.60 0.90 1.23 1.45 1.32 1.58 1.68 1.96 2.02 2.39 2.60 2.57 2.50 2.59 2.47 3.06 2.15 1.97 3.16 2.59 2.66 2.84 2.46 2.83 3.36 2.24 3.69 3.07 0.39 5.30 2.32
0.33 0.38 0.52 0.61 0.62 0.67 0.71 0.74 0.78 0.85 0.92 0.95 0.99 1.05 1.10 1.13 1.13 1.14 1.17 1.18 1.22 1.23 1.25 1.30 1.38 1.45 1.63 1.67 2.18 2.75 1.10
a MUE denotes mean unsigned error (MUE). MMUE ) [MUE(cp) + MUE(no-cp)]/2, and MMMUE ) [MMUE(HB) + MMUE(CT) + MMUE(DI) + MMUE(WI) + MMUE(PPS)]/5; HB, hydrogen bonding; CT, charge transfer; DI, dipole interaction; WI, weak interaction; and PPS, π-π stacking. b We use “no-cp” to denote the calculation without the counterpoise correction for the BSSE and use “cp” to denote the calculation with the counterpoise correction for the BSSE. c The MG3S basis set is used for calculations in this table.
5.4. Noncovalent Interactions. The mean errors for noncovalent interaction are listed in Table 5. In Table 5, we use “no-cp” to denote calculations without the counterpoise correction for the BSSE and we use “cp” to denote calculations that do include the counterpoise correction for the BSSE. In Table 5, we also defined a mean MUE: MMUE ) [MUE(no-cp) + MUE(cp)]/2
(30)
This is a reasonable error criterion because the cp correction is sometimes an overestimate of BSSE and because, in practical work, some calculations are carried out with cp corrections and some without. Table 5 shows that PBE1PBE, M05-2X, PWB6K, and PBE give the best performance for calculating the binding energies of the hydrogen-bonding dimers in the HB6/04 database. Table 5 also shows that M05-2X, PWB6K, and M05 give a very good performance for calculating the binding energies for the complexes in the CT7/04 and DI6/ 04 databases. M05-2X, M05, and PW6B95 give the best performance for calculating the binding energies of the weak interaction complexes in the WI7/05 database. We note that π-π stacking interactions play a dominant role in stabilizing various biopolymers, for example, the double helix structure of DNA, and such interactions are also important for supramolecular design. Table 5 shows that the quality of M05-2X for describing π-π stacking interactions is better than PWB6K. This is encouraging because we have already shown59,60 that PWB6K performs unusually well for
the stacking interactions in the small organic clusters and nucleobase pairs. The overall performance for noncovalent interactions can be judged by the mean MMUE, which is defined as MMMUE ) [MMUE(HB) + MMUE(CT) + MMUE(DI) + MMUE(WI) + MMUE(PPS)]/5 (31) Notice that the five component in eq 31 place different requirements on a density functional. For example, high accuracy for charge-transfer complexes is not well correlated with high accuracy for weak interactions. If we use MMMUE as a criterion to evaluate the overall performance of DFT methods for noncovalent interactions, we can see from Table 5 that M05-2X, PWB6K, M05, MPWB1K, and PW6B95 are the best functionals. 5.5. Composite Results for Main-Group Energetic Databases. Table 6 is a summary of the performance of the tested methods for thermochemistry, kinetics, and noncovalent interactions. The second-to-last column of Table 6 is an average of the three mean unsigned errors. The M05-2X functional has an average error 1.4 times smaller than that of the second best performing method (M05), followed by BMK, MPWB1K, and PW6B95. We also computed a weighted average where each error is divided by the average error of all 30 functionals for that quantity; this is shown in the last column. With this scaled average, the M05-2X functional performs 1.4 times better than the second-best performing functional M05, followed by PW6B95, PWB6K, and MPWB1K.
374 J. Chem. Theory Comput., Vol. 2, No. 2, 2006
Zhao et al.
Table 6. Composite Energetic Results (kcal/mol)
method M05-2X M05 PW6B95 PWB6K MPWB1K MPW1B95 BMK B98 B97-1 B97-3 B1B95 PBE1PBE mPW1PW91 BB1K B97-2 TPSS1KCIS MPW1K B3PW91 B3LYP τ-HCTHh X3LYP TPSSh B1LYP TPSS mPWPW91 BLYP BB95 PBE G96LYP SPWL average
thermononcovalent chemical kinetics interaction scaled TMUE AMUE MMMUE averagea averageb 0.93 0.99 0.81 1.72 1.38 0.98 0.94 1.00 1.07 1.02 0.96 1.31 1.32 1.70 1.02 1.07 2.59 1.14 1.41 1.17 1.89 1.37 2.89 1.36 2.24 1.93 2.40 3.01 2.31 14.70 1.95
0.99 1.44 1.99 1.42 1.35 1.94 1.12 2.57 2.94 1.71 1.79 2.76 2.34 1.27 2.16 3.16 1.32 2.54 3.09 3.13 3.56 4.31 2.73 5.12 5.20 4.90 4.89 6.01 4.26 12.05 3.14
0.33 0.52 0.62 0.38 0.61 0.67 1.10 0.78 0.71 1.18 1.25 0.74 0.95 1.13 1.23 0.99 0.85 1.38 1.13 1.45 0.92 1.05 1.17 1.22 1.30 1.63 1.67 1.14 2.75 2.18 1.10
0.75 0.98 1.14 1.18 1.11 1.20 1.05 1.45 1.57 1.30 1.33 1.60 1.54 1.37 1.47 1.74 1.58 1.69 1.87 1.92 2.12 2.24 2.26 2.57 2.92 2.82 2.99 3.39 3.11 9.64 2.06
0.36 0.48 0.54 0.56 0.56 0.58 0.61 0.68 0.71 0.71 0.73 0.74 0.76 0.77 0.77 0.82 0.84 0.88 0.91 0.97 0.98 1.01 1.14 1.14 1.33 1.35 1.43 1.50 1.68 4.45 1.00
a (TMUE + AMUE + MMMUE)/3 in kcal/mol. b [(TMUE/1.95) + (AMUE/3.14) + (MMMUE/1.10)]/3; note that the scaled average is unitless.
5.6. Trends in Alkyl Bond Dissociation Energies. Recently, Izgorodina et al. reported a study of the performance of several DFT methods, for the prediction of absolute and relative R-X bond dissociation energies (BDEs) where R is an alkyl group (R ) Me, Et, i-Pr, and t-Bu) and X is a substituent (X ) H, CH3, OCH3, OH, and F), and they found that all of the tested DFT methods overestimate the stabilizing effect on BDEs in going from R ) Me to R ) t-Bu, leading in some cases to incorrect qualitative behavior. Note that their results are consistent with the trends for the reaction energies in Table 3 of an earlier paper by Dybala-Defratyka et al.100 Some earlier studies by Curtiss et al.101 had also shown that conventional DFT methods perform much worse for the enthalpies of formation of the larger molecules, and they concluded that this is due to a cumulative effect in the error for the larger molecules. Table 7 summarizes the results for the trends in R-X BDEs (R ) Me and i-Pr; X ) CH3 and OCH3). Table 7 shows that M05-2X gives surprisingly good results for these BDEs; it gives a better performance than the expensive G3RAD102 method for the ABDE4/05 database, and it gives a MUE of only 0.6 kcal/mol, whereas BMK (the second-best DFT method) gives a MUE of 1.7 kcal/mol. Eight functionals were tested against more data of this type, and the results are in Figures 6 and 7 and in the Supporting Information; the additional tests include larger alkyl groups than those present in ABDE4/05, but they yield similar conclusions to those drawn from Table 7. Figures 6 and 7 present the trends for eight relative R-X BDEs (R ) Me, Et, i-Pr, and t-Bu; X ) CH3 and OCH3), and both figures show that M05-2X, like other DFT methods but to a much lesser extent, tends to overestimate the BDE-
Table 7. Alkyl Bond Dissociation Energies (De, kcal/mol)a,b R-CH3 method exptl M05-2X G3-RADc,d BMKd MPW1B95d MPWB1Kd PWB6K B1B95d BB1K PW6B95 B97-1 BB95 PBE B97-2 τ-HCTHh B97-3 B98 X3LYP PBE1PBE M05 mPWPW91 mPW1PW91d B3PW91 MPW1Kd TPSS1KCIS B3LYPd TPSSh TPSS B1LYP BLYPd G96LYP SPWL averagee
R-OCH3
R ) Me R ) i-Pr R ) Me R ) i-Pr 97.39 97.37 96.91 97.99 98.90 98.54 97.96 97.58 97.58 97.26 97.45 98.35 96.79 97.68 96.51 96.76 95.73 95.73 95.23 94.47 94.58 93.28 93.18 92.80 92.11 91.58 90.47 90.48 89.73 90.31 89.01 115.56
95.00 94.01 94.95 93.42 92.78 93.01 92.64 91.05 91.75 91.10 90.83 90.15 89.65 90.18 89.51 89.78 89.10 89.10 89.29 86.99 87.22 87.16 86.52 87.42 85.56 85.01 84.12 83.74 83.44 82.64 80.68 108.51
89.79 90.65 90.53 88.81 88.79 88.10 87.48 87.83 87.02 87.28 87.05 87.79 87.24 86.77 86.93 85.86 86.06 86.06 85.63 86.32 85.26 84.37 83.79 82.77 83.07 82.58 82.08 82.36 80.46 81.09 79.64 108.10
91.51 90.93 92.87 87.99 86.68 86.56 86.21 86.01 85.17 85.18 84.29 83.69 84.08 83.10 83.78 82.78 83.31 83.31 83.59 82.77 81.87 82.86 81.07 81.25 80.44 80.06 79.62 79.54 78.25 77.50 75.40 105.49
MSE
MUE
-0.18 0.39 -1.37 -1.64 -1.87 -2.35 -2.80 -3.04 -3.22 -3.52 -3.43 -3.98 -3.99 -4.24 -4.63 -4.87 -4.87 -4.98 -5.79 -6.19 -6.51 -7.28 -7.36 -8.13 -8.62 -9.35 -9.39 -10.45 -10.53 -12.24 15.99
0.61 0.66 1.67 2.39 2.44 2.64 2.90 3.14 3.22 3.55 3.91 3.98 4.14 4.24 4.63 4.87 4.87 4.98 5.79 6.19 6.51 7.28 7.36 8.13 8.62 9.35 9.39 10.45 10.53 12.24 15.99 5.70
a The B3LYP/6-31G(d) geometries are used in all calculations in this table. b All DFT calculations in this table use the 6-311+G(3df,2p) basis set. c G3-RAD is the “Gaussian-3 for radicals” method of ref 102. d Data for these methods are taken from Izgorodina et al.71 e Average excludes G3-RAD, which is a wave function method (not a density functional method).
Figure 6. Effect of level of theory on the relative bond dissociation energies (in kcal/mol) for R-CH3 species (R ) methyl, ethyl, isopropyl, tert-butyl).
lowering effect accompanying the increasing size of the alkyl group. In contrast, the wave-function-based method G3-RAD slightly underestimates the stabilizing effect on R-X BDEs
Design of Density Functionals
J. Chem. Theory Comput., Vol. 2, No. 2, 2006 375 Table 8. MUE (kcal/mol) for the TMAE4/05 and MLBE4/ 05 Databases with the DZQ Basis Set TMAE4/05
Figure 7. Effect of level of theory on the relative bond dissociation energies (in kcal/mol) for R-OCH3 species (R ) methyl, ethyl, isopropyl, tert-butyl).
on going from R ) Me to t-Bu. Since Table 7 shows smaller mean unsigned errors for M05-2X than for G3-RAD, it is no longer appropriate to consider this kind of error as a failure of DFT, although it is a failure of some functionals. The results in Table 7 and Figures 6 and 7 are encouraging because M05-2X shows small errors for the absolute and relatiVe BDEs, and M05-2X offers promise as a reliable functional for larger systems. 5.7. Transition-Metal-Transition-Metal and MetalLigand Bond Energies. Metal-metal and metal-ligand bonding is very important in many application areas.103-115 Table 8 summarizes the results for the TMAE4/05 and MLBE4/05 databases. For the TMAE4/05 database of bond energies of transition-metal dimers, BLYP, G96LYP, PBE, mPWPW91, and M05 give the best results. Note that M052X, PWB6K, MPWB1K, BB1K, and BMK are among the worst methods for transition-metal dimers because these DFT methods contain a large amount of HF exchange, and this makes the functionals less valid for systems with significant nondynamical correlation energy; hence, methods with correlation functionals that primarily account for dynamical (not static) correlation (this includes all 30 functionals tested in this article) and with more than 30% HF exchange are not recommended for studies of the interactions of transitionmetal atoms with other transition-metal atoms where nondynamical correlation plays an important role; we will come back to this point in section 5.12. For the MLBE4/05 database of metal-ligand compounds, TPSS1KCIS, TPSSh, M05, B97-2, and PBE1PBE give the best performance. In Table 8, MMUE is the average of the MUE for the TMAE4/05 and MLBE4/05 databases, and BLYP, M05, G96LYP, TPSS, and B97-2 give the smallest MMUEs. Notice that, of the 11 functionals with the smallest MMUEs, only M05 (X ) 28) and B97-2 (X ) 21) have X values larger than 15; six of these functionals have X ) 0, and three (TPSSh, τ-HCTHh, and TPSS1KCIS) have X values in the range 10-15. The ability to obtain good results
MLBE4/05
method
MSE
MUE
MSE
MUE
MMUEa
BLYP M05 G96LYP TPSS B97-2 mPWPW91 PBE BB95 TPSSh τ-HCTHh TPSS1KCIS B97-1 B3LYP B98 X3LYP B3PW91 PBE1PBE PW6B95 B1B95 MPW1B95 B97-3 mPW1PW91 MPWB1K BB1K B1LYP M05-2X MPW1K PWB6K SPWL BMK average
-0.86 -5.98 -5.71 -6.18 -10.07 -4.03 0.38 3.32 -15.97 -5.91 -18.26 -17.70 -21.47 -19.67 -21.10 -25.34 -25.04 -24.32 -25.13 -25.06 -22.80 -26.46 -29.30 -29.56 -27.14 -21.92 -31.83 -33.90 23.03 -35.98
1.97 7.34 5.71 8.38 10.95 7.28 5.87 7.98 15.97 13.07 18.26 18.64 21.47 19.92 21.10 25.34 25.04 24.32 25.13 25.06 22.80 26.46 29.30 29.56 27.14 29.42 31.83 33.90 23.03 36.81 19.97
9.23 -2.20 6.99 7.00 -0.61 9.89 12.12 12.13 1.42 3.56 0.79 0.67 -1.28 -0.73 -1.75 -2.54 -3.34 -4.00 -4.40 -4.64 -4.98 -4.72 -11.35 -11.04 -16.72 -12.18 -13.10 -13.63 30.30 13.35
9.23 4.97 8.10 7.00 5.52 9.89 12.12 12.13 4.62 7.68 4.39 8.36 6.44 8.00 6.86 5.46 6.31 7.36 7.16 7.61 10.52 7.04 11.52 11.66 16.92 15.24 13.10 13.63 30.30 17.74 9.90
5.60 6.15 6.90 7.69 8.24 8.58 9.00 10.05 10.30 10.38 11.32 13.50 13.95 13.96 13.98 15.40 15.68 15.84 16.15 16.33 16.66 16.75 20.41 20.61 22.03 22.33 22.46 23.77 26.66 27.27 14.93
a
MMUE ) [MUE(TMAE4/05) + MUE(MLBE4/05)]/2.
for bonds to metal atoms with an X value as large as 28 is one of the characteristics that allows M05 to have a broader range of applicability than any other functional. 5.8. Tests for Dipole Moments. Table 9 presents the performance for the DM6/05 database of the dipole moments. NH2(CHdCH)6NO2 (denoted as N6) is a push-pull π-conjugated system, and the accurate evaluation of the electric dipole properties for this type of molecule is a difficult problem for density functional theory.75,76 Among the tested DFT methods, M05-2X gives the best results for the dipole moment of N6, and in general, the DFT methods with higher percentages of Hartree-Fock exchange perform better than the DFT methods with lower (or zero) percentages of Hartree-Fock exchange. Overall, PWB6K gives the lowest MUE, followed by M05-2X, MPWB1K, BB1K, and BMK. If we consider the MU%E, MPW1B95 give the lowest MU%E, followed by PWB6K, M05, MPWB1K, and BB1K. 5.9. Tests for Noncovalent Interactions in Nucleobase Pairs. Table 10 summarizes the results for the stacking and hydrogen-bonding interactions in nucleobase pairs. All of the structures have been detailed in a previous paper,60 and they are also given in the Supporting Information. For stacking interactions, SPWL and M05-2X give the best performance. However, the good performance of SPWL for stacking interactions is not matched by good accuracy for hydrogen bonding. Table 9 shows that SPWL gives the largest errors for the hydrogen-bonding interactions, while M052X gives the best performance for the interaction energies of the two Watson-Crick hydrogen-bonded base pairs. The average MUE in Table 10 is defined as AMUE ) MUE(stacking) + MUE(hydrogen bonding) (32)
376 J. Chem. Theory Comput., Vol. 2, No. 2, 2006
Zhao et al.
Table 9. Dipole Moments Predicted by Density Functionalsa,b method
N6c
LiCl
H2CO
CuH
H2O
BF
accurate PWB6K M05-2X MPWB1K BB1K BMK MPW1K MPW1B95 B1B95 PW6B95 B97-3 PBE1PBE mPW1PW91 M05 B97-2 B98 B97-1 B1LYP B3PW91 X3LYP B3LYP TPSS1KCIS τ-HCTHh TPSSh TPSS mPWPW91 BB95 PBE G96LYP BLYP SPWL average
11.56f
7.23g
2.39g
2.97h
1.85g
0.79g
15.00 14.85 15.07 15.12 15.27 15.20 15.55 15.66 15.68 15.76 15.85 15.87 15.97 16.04 16.07 16.06 15.94 16.11 16.10 16.18 16.38 16.41 16.54 17.06 17.10 17.04 17.08 17.18 17.17 17.42
7.19 7.16 7.19 7.20 7.26 7.21 7.13 7.14 7.11 7.17 7.11 7.13 7.08 7.14 7.11 7.10 7.08 7.12 7.06 7.07 7.08 7.11 7.09 7.03 6.98 6.98 6.96 7.01 6.94 6.95
2.59 2.68 2.58 2.56 2.63 2.60 2.49 2.47 2.48 2.50 2.47 2.49 2.52 2.46 2.48 2.46 2.51 2.46 2.49 2.48 2.42 2.44 2.41 2.34 2.31 2.28 2.29 2.34 2.33 2.37
3.23 3.25 3.20 3.15 2.91 3.25 2.92 2.89 2.87 2.90 2.88 2.89 2.91 2.91 2.84 2.80 2.77 2.79 2.70 2.68 2.76 2.72 2.81 2.60 2.36 2.28 2.34 2.29 2.25 2.16
2.01 2.04 2.01 2.00 2.04 2.01 1.98 1.98 1.98 1.98 1.97 1.98 2.00 1.96 1.97 1.96 1.98 1.97 1.98 1.98 1.95 1.96 1.94 1.92 1.93 1.92 1.93 1.94 1.93 2.00
0.80 0.81 0.81 0.83 0.87 0.86 0.85 0.87 0.84 0.92 0.92 0.91 0.80 0.91 0.91 0.92 0.88 0.93 0.89 0.90 0.94 0.92 0.97 1.01 0.98 0.96 1.00 0.99 0.96 0.99
MS%Ed
MU%Ee
MSE
MUE
9.2 10.2 9.3 9.4 10.2 11.1 8.2 8.5 7.8 10.2 9.8 9.7 8.0 9.9 9.7 9.5 8.7 9.8 8.6 8.6 9.5 9.4 10.7 10.2 8.1 6.9 8.2 8.3 7.3 8.6
9.4 10.6 9.5 9.5 10.9 11.2 9.3 9.9 9.6 11.3 11.4 11.2 9.5 11.1 11.8 12.0 11.7 12.3 12.4 12.7 12.6 12.7 13.2 16.0 17.4 17.4 17.9 17.8 17.6 19.4 12.6
0.67 0.66 0.68 0.68 0.70 0.72 0.69 0.70 0.69 0.74 0.74 0.74 0.75 0.77 0.76 0.75 0.73 0.76 0.74 0.75 0.79 0.80 0.83 0.86 0.81 0.78 0.80 0.83 0.80 0.85
0.69 0.69 0.69 0.69 0.72 0.73 0.74 0.76 0.77 0.79 0.81 0.81 0.82 0.82 0.85 0.85 0.85 0.86 0.89 0.90 0.91 0.92 0.93 1.07 1.13 1.13 1.13 1.15 1.16 1.22 0.88
a All values are in Debyes. b All DFT calculations are single-point calculations using the TZQ basis set. c NH (CHdCH) NO is denoted as 2 6 2 N6. d Mean percentage signed error. e Mean percentage unsigned error. f MP2/6-311+G(2df,2p) result. All calculations use MP2/6-31G geometry for this molecule. g CCSD(T)/aug-cc-pVTZ result. All calculations use CCSD(T)/aug-cc-pVTZ geometry for these molecules. h The reference dipole moment for CuH is an average of the values by the MCPF calculation and a CCSD(T)/ANO calculation performed in the present study, where ANO is the triple-ζ atomic natural orbital basis set of Widmark et al. The geometry is taken from a previous study by Langhoff and Bauschlicher, and all calculations use this geometry (rCu-H ) 1.509 Å).
Table 10. Results for Stacking and Hydrogen-Bonding Interactions in Nucleobase Pairs (kcal/mol) stacking methods estimatec
best M05-2X PWB6K MPWB1K PW6B95 MPW1B95 M05 PBE1PBE B97-1 BMK τ-HCTHh TPSSh SPWL B3LYP B97-3 average
A‚‚‚T S G‚‚‚C S C‚‚‚C 11.60 10.28 9.50 8.19 7.68 7.47 5.77 3.54 3.54 5.42 2.39 1.42 12.59 -0.10 0.71
16.90 16.25 14.86 13.68 13.10 12.83 11.95 10.44 10.26 11.54 9.38 8.47 18.90 7.39 7.60
APa
9.90 10.52 10.88 9.63 9.48 8.98 7.86 5.30 5.64 6.37 4.54 3.75 11.62 2.87 3.45
C‚‚‚C
Da
9.43 10.02 9.66 8.63 8.46 8.08 7.66 5.81 6.31 6.17 5.14 4.46 10.70 3.64 4.24
hydrogen bonding C‚‚‚C
Sa
-2.45 -5.08 -5.93 -7.02 -6.56 -7.10 -6.40 -8.40 -8.01 -9.14 -8.87 -9.64 -2.33 -10.70 -10.21
U‚‚‚U S
MSE
MUE A‚‚‚T WC G‚‚‚C WC
10.30 8.76 7.94 6.51 6.45 6.01 5.79 3.97 4.05 4.42 3.05 2.29 10.64 1.47 1.90
-0.82 -1.46 -2.68 -2.84 -3.24 -3.84 -5.84 -5.65 -5.15 -6.68 -7.49 1.07 -8.52 -8.00
1.22 1.86 2.68 2.84 3.24 3.84 5.84 5.65 5.15 6.68 7.49 1.07 8.52 8.00 4.58
15.40 14.56 14.22 13.42 13.26 13.18 13.68 14.42 14.08 12.49 13.79 13.37 22.30 12.73 11.78
28.80 28.58 28.39 27.45 26.68 26.80 27.07 28.43 27.44 26.30 27.42 26.80 39.44 26.17 24.78
MSE
MUE AMUEb
-0.53 -0.79 -1.67 -2.13 -2.11 -1.72 -0.67 -1.34 -2.71 -1.49 -2.02 8.77 -2.65 -3.82
0.53 0.79 1.67 2.13 2.11 1.72 0.67 1.34 2.71 1.49 2.02 8.77 2.65 3.82 2.32
0.87 1.33 2.17 2.49 2.67 2.78 3.25 3.50 3.93 4.08 4.75 4.92 5.59 5.91 3.45
a 6-31+G(d,p) is used for all calculations in this table. b AP denotes antiparallel, D denotes displaced, and S denotes sandwich. The structures for all base pairs in this table can be found in ref 60 and in the Supporting Information. c See ref 60 and references therein for the sources of these best estimates. d AMUE ) 0.5MUE(Stacking) + 0.5MUE(Hydrogen Bonding).
M05-2X gives the lowest AMUE, followed by PWB6K and MPWB1K. 5.10. Tests for One-Electron Systems. Table 11 presents the results for three one-electron systems, namely, the hydrogen atom, H2+ with a bond distance of 1.4 bohr, and H2+ with a bond distance of 2.0 bohr. In Table 11, PWB6K gives the lowest MUE, followed by BB1K and BMK. All of the mean errors are disconcertingly large, but it is encouraging that the better functionals have errors 1.5-3 times smaller than those of the popular B3LYP. Again, the
DFT methods with higher percentages of Hartree-Fock exchange generally (but not always) perform better than the DFT methods with lower (or zero) percentages of HartreeFock exchange. It is interesting to note that five of the six best functionals for thermochemical kinetics (Table 4) are also among the six best density functionals in Table 11. B1B95 and MPW1B95 also rank in the top 10 of both tables. However, one cannot generalize this result because there are also cases where the performance in these two tables does not correlate.
Design of Density Functionals
J. Chem. Theory Comput., Vol. 2, No. 2, 2006 377
Table 11. Predicted Energy for H and H2+a energy (hartree)
mean errors (kcal/mol)
+
+
methods
H
H2 (1.4 b)
H2 (2.0 b)
MSE
MUE
HF PWB6K BB1K BMK X3LYP MPWB1K M05-2X B1B95 B1LYP PBE MPW1B95 M05 G96LYP BB95 BLYP PW6B95 TPSS1KCIS PBE1PBE TPSSh TPSS B3LYP B98 mPWPW91 B97-1 mPW1PW91 MPW1K B3PW91 B97-3 B97-2 τ-HCTHh SPWL average
-0.499946 -0.500452 -0.498539 -0.498903 -0.499785 -0.497995 -0.499743 -0.498260 -0.498204 -0.499854 -0.497603 -0.497839 -0.499052 -0.497781 -0.497781 -0.501499 -0.500036 -0.501227 -0.500043 -0.500069 -0.502346 -0.502865 -0.503098 -0.502785 -0.503839 -0.504420 -0.504154 -0.503829 -0.504206 -0.507268 -0.478593
-0.569830 -0.570071 -0.568421 -0.567576 -0.569126 -0.567491 -0.571127 -0.568153 -0.568099 -0.569849 -0.567017 -0.570464 -0.570323 -0.567708 -0.567708 -0.571477 -0.572567 -0.571595 -0.572672 -0.573028 -0.572079 -0.574646 -0.574019 -0.574955 -0.574884 -0.575563 -0.575088 -0.575986 -0.578058 -0.580641 -0.540711
-0.602521 -0.605866 -0.604787 -0.604547 -0.607086 -0.603772 -0.607252 -0.605486 -0.605644 -0.609222 -0.604200 -0.607749 -0.609532 -0.607079 -0.607079 -0.608338 -0.609298 -0.609083 -0.609564 -0.610440 -0.610047 -0.612113 -0.612966 -0.612360 -0.612020 -0.611473 -0.612744 -0.613022 -0.615081 -0.618280 -0.583762
0b -0.86 0.11 0.27 -0.77 0.64 -1.22 0.08 0.07 -1.39 0.73 -0.79 -1.38 -0.06 -0.06 -1.89 -2.01 -2.01 -2.09 -2.35 -2.55 -3.62 -3.72 -3.72 -3.86 -4.01 -4.12 -4.30 -5.24 -7.09 14.48
0b 0.86 1.06 1.11 1.14 1.16 1.30 1.32 1.38 1.42 1.43 1.67 1.76 1.85 1.85 1.89 2.01 2.01 2.09 2.35 2.55 3.62 3.72 3.72 3.86 4.01 4.12 4.30 5.24 7.09 14.48 2.88
a The cc-pVQZ basis set is employed in all calculations in this table. b For a one-electron system, Hartree-Fock is the same as full configuration interaction for a given basis set, and the error in the density functional calculations are computed relative to these results.
For example, MPW1K and B97-3 rank much higher in Table 4 than in Table 11, and X3LYP and PBE rank much higher in Table 11 than in Table 4. 5.11. Tests for a Donor-Acceptor System: HCN-BF3. Recently, Philips and Cramer116 reported a study of a boronnitrogen complex, namely, HCN-BF3.117,118 This is an example of a Lewis acid-base complex, also called a dative bond or a coordinate covalent bond. They employed 12 GGA and hybrid GGA functionals as well as some wave-functionbased methods, and their conclusion was that “all DFT methods fail to predict a binding energy that compares favorably to the MCG3//MC-QCISD result of -5.7 kcal/ mol.” In particular, all DFT methods tested gave bond energies in the range 1.8-4.3 kcal/mol, except MPW1K and BLYP, which respectively yielded 4.7 and 7.4 kcal/mol. Table 12 shows the results for the 14 DFT methods tested in Table 10; of these, 13 were not tested in Philips and Cramer’s paper. Table 12 also includes the B3PW91 method, which was judged116 the overall best for structures and frequencies; MPW1K, which was the best (of those functionals tested116) for complexation energy; and PBE, TPSS, BB95, B1B95, and BB1K, which are added for their fundamental interest and to illustrate the dependence on the fraction of Hartree-Fock exchange. Table 12 shows that the binding energies calculated by the PWB6K, MPWB1K, and M05-2X methods agree well with the best estimate (MCG3// MC-QCISD/3 calculation), and M05-2X even predicts more attraction than MCG3//MC-QCISD/3. It is encouraging that M05 is the most successful functional for this system and that Table 12 shows that almost all of the functionals in Table 10 are better on average than B3PW91; M05 and PW6B95
Table 12. Bond Length, Dipole Moment, and Binding Energy of HCN-BF3a methods
RBN (Å)
µ (D)
∆E (kcal/mol)
M%UEb
best estimate M05 PW6B95 MP2 B97-1 PWB6K MPW1B95 M05-2X PBE1PBE MPWB1K PBE τ-HCTHh B3LYP MPW1K BB1K BMK B1B95 B97-3 B3PW91 TPSSh TPSS BB95 SPWL averagef
2.473c 2.492 2.427 2.361 2.500 2.292 2.348 2.352 2.348 2.253 2.407 2.426 2.535 2.323 2.346 2.351 2.432 2.615 2.465 2.230 2.239 2.538 1.731 2.364
4.14d 4.30 4.39 4.51 4.29 4.77 4.60 4.73 4.64 4.89 4.45 4.46 4.23 4.74 4.63 4.72 4.39 4.09 4.37 4.98 4.93 4.12 6.97 4.65
-5.7e -5.0 -5.0 -6.3 -4.6 -5.9 -4.9 -6.5 -4.8 -5.3 -4.3 -4.2 -3.8 -4.7 -4.5 -4.5 -3.8 -3.6 -3.2 -4.3 -4.1 -2.7 -12.1 -4.9
5.6 6.5 8.0 8.1 8.7 10.2 11.3 11.1 11.3 11.4 11.9 12.7 12.7 12.9 13.3 13.6 14.6 16.6 18.1 18.8 18.6 70.4 14.8
a All DFT and MP2 results are for the aug-cc-pVTZ basis set. Mean percentage unsigned error. c Experimental result.117 d Experimental result.118 e MCG3//MCQCISD/3 result.116 f Average excludes MP2, which is a wave function method, not a density functional method.
b
are more accurate, on average, by a factor of 2.9 and 2.5, respectively, as well as being more accurate than MP2. M052X reduces the error in B3PW91, on average, by 33%. 5.12. Multireference Character. A simple and useful way to describe the optimum domains of applicability of the M05 and M05-2X functionals is that the former is recommended
378 J. Chem. Theory Comput., Vol. 2, No. 2, 2006
Zhao et al.
for systems containing metals or transition metals (especially those in groups 2-10) and the latter is recommended for systems containing only nonmetallic or only main-group elements (although M05 is also very good for such systems, as shown in Tables 6, 8, and 10). This way of classifying systems, though, does not really capture the essence of the issue at a higher level of sophistication. We believe that the essential distinction is multireference character. Systems with significant multireference character are not well described by most density functionals that have more than 5-15% Hartree-Fock exchange. A system with large multireference character is one for which no single configuration-state function provides a good zero-order description;119,120 such a system is said to contain significant amounts of static, neardegeneracy, or nondynamical correlation energy, often associated with multicenter systems, but also found in atoms.121 Having made this distinction, one might summarize the situation as follows: M05 is recommended for applications where the systems studied involve both multireference and single-reference behavior, whereas if only single-reference behavior is to be encountered, one can obtain higher quantitative accuracy by switching to M05-2X. This is more satisfactory than the formulation at the start of this paragraph, but only partly more satisfactory because “multireference character” is not completely unambiguous. One can characterize multireference character by analyzing a configuration interaction122 or coupled cluster calculation,123 but this is often impractical. In a recent paper,66 we proposed a simpler criterion for the multireference character of a bond. We called this the B1 diagnostic and defined it as B1 ) [De(BLYP) - De(B1LYP//BLYP)]/n
(33)
where De is the energy required to break n bonds and B1LYP//BLYP denotes a B1LYP calculation of the same quantity using the BLYP equilibrium geometries for the molecule and the fragments. For B1LYP, the percentage of Hartree-Fock exchange is 25. The B1 diagnostic measures multireference character because the Hartree-Fock exchange approximation fails badly for multireference systems, whereas GGAs can usually handle these systems almost as well as they handle single-reference systems. We previously concluded that bonds with B1 j 10 kcal/mol are reasonably classified as single-reference cases, whereas those with B1 J 10 kcal/mol should be classified as multireference. This criterion is clearly not sophisticated enough to supplant system-specific discussions of metallic and multireference character,77,120,121,124-133 and it does not fully supersede characterizing bonds in chemical terms,66 but its ease of use is appealing. In Tables 13 and 14, we present results for eight systems, four of which have B1 values less than 10 kcal/mol (single reference) and four of which have B1 values greater than 10 kcal/mol (multireference). Ozone (O3) is a well-studied multireference system,122 and its B1 is about 22 kcal/mol. The cases in Table 13 were chosen so that two of the singlereference cases involve transition metals, one involves a main-group metal, and one has no metals. Similarly, two of the multireference cases involve metallic elements and two involve only nonmetals. Table 14 shows that the quality of
Table 13. Dissociation Energies (kcal/mol) and B1 Values for Eight Bond-Breaking Processesa De process
experimentb
BLYP
B1LYP//BLYP
B1
CH4 f CH3 + H LiCl f Li + Cl AgCu f Ag + Cu Cu2 f 2Cu VS f V + S CN f C + N O3 f 3O ZrV f Zr + V
112.7 113.9 40.9 47.2 106.9 180.6 146.1 61.9
109.9 108.2 41.7 46.4 111.1 190.9 170.1 72.7
109.5 107.4 36.2 39.7 93.7 173.0 126.3 32.9
0.4 0.8 5.5 6.7 17.4 17.9 21.9c 39.8
a The TZQ basis set is used. The TZQ basis always uses spherical harmonic d and f functions (5D 7F sets). b The experimental values for CH3-H and CN are calculated by using the experimental atomization energies from Database/3.70 The experimental value for O3 is taken from Database4/05.73 The experimental values for Cu2 and AgCu are taken from a previous paper,57 and the experimental values for VS and ZrV are taken from a recent paper.66 All dissociation energies in this table are zero-point-exclusive and spin-orbit-inclusive. c We put n ) 2 in eq 33 in this case because two bonds are broken (not counting the long “bond”).
the predictions depends more on multireference character than on the metallic character. For the prediction of the bond energies in systems with a low B1 value, M05 and M05-2X perform equally well, but for the systems with a high B1 value, M05-2X performs much worse than M05. Note that eq 33 does not apply to transition states, but the reader should be aware that transition states, even those for radical reactions, are not all multireference systems, although it is a common misconception that they are. For example, a multireference plus single and double excitation calculation lowers the barrier height of the H + H2 reactions by only 0.3 kcal/mol as compared to a single-reference calculation with single and double excitations.134 Similarly, single-reference plus dynamical-correlation-energy treatments give reasonable descriptions of the F-H-H and H-F-H transition states.135 These conclusions based on wave function theory are consistent with our DFT findings that several methods fail quite badly for multireference systems with B1 > 10 kcal/ mol but are nevertheless quite accurate for transition states, even radical transition states. Examples would be MPW1K, BB1K, and PWB6K. In this light, the good performance of the M05 functional both for B1 > 10 kcal/mol systems and for barrier heights is even more dramatic. A final comment on transition-metal systems is warranted. In particular, it should be noted that complexes in which a transition metal is saturated with ligands or is one ligand short of saturation may have far less multireference character than highly unsaturated systems such as metal-containing diatomic molecules. Thus, the cases in our metallic training sets are more difficult than the kinds of transition-metal complexes that occur in many areas of organometallic chemistry and metalloenzyme chemistry.110,111 Nevertheless, there are many other important applications where the valence state, magnetic state, or oxidation state of the metal is unknown or changes during a reaction, and the results in Table 13 and the high-B1 section of Table 14 provide an indication of the ability of various density functionals to treat this important class of problems. 5.13. Self-Exchange. One of the key sources of error in density functional theory is self-exchange.136-138 For ex-
Design of Density Functionals
J. Chem. Theory Comput., Vol. 2, No. 2, 2006 379
Table 14. Signed Errors and Mean Unsigned Errors (kcal/mol) in Bond Energiesa B1 < 10
B1 > 10
methods
Xb
Cu2
AgCu
CH3-H
LiCl
MUE
ZrV
VS
O3
CN
MUE
MMUEc
M05 B3LYP B1B95 BLYP PBE1PBE mPWPW91 mPW1PW91 B1LYP PBE M05-2X BB95 BB1K average
28 20 28 0 25 0 25 25 0 56 0 42
0.0 -5.6 -5.5 -0.8 -6.3 -0.5 -7.8 -7.5 1.4 0.8 2.2 -8.6
1.2 -2.9 -2.6 0.8 -3.2 1.5 -4.5 -4.9 3.2 3.6 3.7 -5.1
-1.7 -1.7 -0.5 -2.8 -3.3 -3.1 -3.7 -3.2 -2.6 -0.7 -0.2 -0.6
-2.6 -4.9 -5.1 -5.7 -5.4 -4.7 -5.7 -6.5 -4.3 1.7 -2.4 -5.2
1.4 3.8 3.4 2.5 4.6 2.4 5.4 5.5 2.9 1.7 2.1 4.9 3.4
-12.2 -19.1 -23.7 10.8 -25.1 9.0 -28.3 -25.6 13.4 -42.7 20.3 -41.2
-2.7 -8.6 -10.1 4.2 -11.1 4.9 -12.7 -11.4 7.0 -17.8 9.1 -17.3
-7.2 -5.8 -5.9 24.0 -6.1 32.4 -11.3 -19.7 38.6 -20.6 41.7 -27.3
-1.1 -1.5 -3.5 10.3 -2.0 12.9 -4.8 -7.6 16.6 -7.0 16.1 -12.4
5.8 8.8 10.8 12.3 11.1 14.8 14.3 16.1 18.9 22.0 21.8 24.6
3.6 6.3 7.1 7.4 7.8 8.6 9.8 10.8 10.9 11.9 12.0 14.7 9.2
a All DFT calculations in this table use the TZQ basis set with consistently optimized geometries. b Percentage of Hartree-Fock exchange in each functional. c MMUE ) 0.5[MUE(B1 < 10) + MUE(B1 > 10)].
ample, self-exchange is responsible for the poor performance of time-dependent DFT for charge-transfer excited states.139 One direction of some current research in DFT is to try to obtain correlation functionals that perform well even with 100% Hartree exchange,38,62,63 which eliminates the selfexchange problem. The present functional does not achieve this goal, but it does perform well with 56% Hartree-Fock exchange, which is much higher than the fraction of HartreeFock exchange, 20-28%7,8,10,15,16,19,22,32,47,56,58,64,65 or, in one case, 31%,51 of previous functionals with good generalpurpose performance and is even higher than the fraction of Hartree-Fock exchange, 42-46%,27,49-51,58 of functionals designed especially for chemical kinetics. And yet, the M052X functional gives better performance than any of the functionals for thermochemical kinetics and alkyl bond energies. Thus, the M05-2X functional should ameliorate some of the problems caused by spurious self-exchange. Furthermore, both M05-2X and M05 are completely free of self-correlation error. 5.14. Comment on Functional Development. A lesson reinforced by the present work is that a good training set is very helpful in parametrizing density functionals, but it is not sufficient. The previous functional forms, prior to M05, did not take full advantage of kinetic energy density and its combination with constraint satisfaction, and they are unable to provide the kind of performance we achieved with M05 and M05-2X. Designing the dependence of the exchangecorrelation functional on kinetic energy density was the key to the improved performance achieved here, as compared (for example) to our previous PW6B95 and PWB6K functionals. In designing the new functional form, we built on several key insights in the work of Becke,10,18,25 but we combined them in new ways and extended them to allow greater flexibility while satisfying the uniform electron gas limit and self-correlation-free limits. In addition, we simultaneously optimized the correlation functional, the exchange functional, and the fraction of Hartree-Fock exchange. It is well-known that it is important for the exchange and correlation functionals to be well-matched. This is partly because they separately have the wrong form at long range and also because exchange density functionals include not only exchange but also near-degeneracy correlation,7,25,29,140-143 whereas the correlation functional includes only dynamical correlation. It is important to balance the inclusion of near-
degeneracy correlation, which is necessary to treat multireference character, with Hartree-Fock exchange, which eliminates (or partially eliminates, when X < 100 in eq 20) spurious self-exchange interactions.
6. Concluding Remarks This paper presents a new hybrid meta exchange-correlation functional, M05-2X, for thermochemistry, thermochemical kinetics, and noncovalent interactions. It also presents a more complete picture of the original M05 functional that was originally defined in a preliminary communication. These two functionals incorporate kinetic energy density in a balanced way in the exchange and correlation functionals; they satisfy the uniform electron gas limit, and they are selfcorrelation-free. They were comparatively assessed against the MGAE109/3 main-group atomization energy database; the IP13/3 ionization potential database; the EA13/3 electron affinity database; the HTBH38/4 database of barrier heights for hydrogen-transfer reactions; the HB6/04 hydrogenbonding database; the CT7/04 charge-transfer database; the DI6/04 dipole interaction database; the WI7/05 weak interaction database; the PPS5/05 π-π stacking database; the ABDE4/05 alkyl bond dissociation energy database; the TMAE4/05 database for transition-metal dimers; the MLBE4/ 05 database for metal-ligand compounds; a dipole moment database, DM6/05, and accurate results for nucleobase interaction energies; the absolute energies of one-electron systems; and the properties of a Lewis acid-base complex, HCN-BF3. From these assessments and from a comparison to results for 28 functionals in the literature, we draw the following conclusions, based on an analysis of mean unsigned errors: (1) The M05-2X, M05, PW6B95, PWB6K, and MPWB1K functionals give the best results for a combination of nonmetallic thermochemical kinetics, thermochemistry, and noncovalent interactions. (2) The M052X method gives the best performance for the calculation of absolute and relative bond dissociation energies for singlereference systems and for calculations of noncovalent interactions between nucleobases. (3) The M05 functional gives, in addition, good performance for multireference systems, including metals. From the present study, we recommend M05-2X, M05, PW6B95, PWB6K, and MPWB1K for general purpose applications in thermochemistry and kinetics, and we espe-
380 J. Chem. Theory Comput., Vol. 2, No. 2, 2006
Zhao et al.
cially recommend M05-2X for calculating bond dissociation energies. For systems involving transition-metal bonding and other multireference systems, we recommend the M05 functional. It is very encouraging that we succeeded in developing density functionals with very broad applicability. They should be especially useful for many applications in chemistry and for condensed-phase systems and molecular recognition problems (including supramolecular chemistry and protein assemblies) where noncovalent interactions are very important.
(18) Becke, A. D. J. Chem. Phys. 1998, 109, 2092.
Acknowledgment. We are grateful to Jan Martin and Mark Iron for assistance with the BMK functional and to Benoit Champagne for supplying the N6 geometry. This work was supported in part by the U.S. Department of Energy, Office of Basic Energy Sciences, by the National Science Foundation, and by the Office of Naval Research.
(24) Adamo, C.; Ernzerhof, M.; Scuseria, G. E. J. Chem. Phys. 2000, 112, 2643.
Note Added after ASAP Publication. This article was released ASAP on February 4, 2006, with an agency that helped support this work not included in the Acknowledgment and with minor errors in Table 13. The correct version was posted on March 7, 2006. Supporting Information Available: All of the databases are given. The geometries of the nucleobase pairs, of the DM6/05 database, and of the noncovalent interactions are also given. This material is available free of charge via the Internet at http://pubs.acs.org. References (1) Becke, A. D. J. Chem. Phys. 1986, 84, 4524.
(19) Hamprecht, F. A.; Cohen, A. J.; Tozer, D. J.; Handy, N. C. J. Chem. Phys. 1998, 109, 6264. (20) Handy, N. C.; Tozer, D. J. Mol. Phys. 1998, 94, 707. (21) Krieger, J. B.; Chen, J.; Iafrate, G. J.; Savin, A. In Electron Correlations and Materials Properties; Gonis, A., Kioussis, N., Eds.; Plenum: New York, 1999; p 463. (22) Adamo, C.; Barone, V. J. Chem. Phys. 1999, 110, 6158. (23) Perdew, J. P.; Kurth, S.; Zupan, A.; Blaha, P. Phys. ReV. Lett. 1999, 82, 2544.
(25) Becke, A. D. J. Chem. Phys. 2000, 112, 4020. (26) Rabuck, A. D.; Scuseria, G. E. Theor. Chem. Acc. 2000, 104, 439. (27) Lynch, B. J.; Fast, P. L.; Harris, M.; Truhlar, D. G. J. Phys. Chem. A 2000, 104, 4811. (28) Proynov, E.; Chermette, H.; Salahub, D. R. J. Chem. Phys. 2000, 113, 10013. (29) Handy, N. C.; Cohen, A. J. Mol. Phys. 2001, 99, 403. (30) Perdew, J. P.; Schmidt, K. In Density Functional Theory and Its Applications to Materials; Van-Doren, V., Alsenoy, C. V., Geerlings., P., Eds.; American Institute of Physics: New York, 2001. (31) Lynch, B. J.; Truhlar, D. G. J. Phys. Chem. A 2001, 105, 2936. (32) Wilson, P. J.; Bradley, T. J.; Tozer, D. J. J. Chem. Phys. 2001, 115, 9233. (33) Parthiban, S.; de Oliveira, G.; Martin, J. M. L. J. Phys. Chem. A 2001, 105, 895. (34) Boese, A. D.; Handy, N. C. J. Chem. Phys. 2002, 116, 9559.
(2) Becke, A. D. Phys. ReV. A: At., Mol., Opt. Phys. 1988, 38, 3098.
(35) Baker, J.; Pulay, P. J. Chem. Phys. 2002, 117, 1441.
(3) Lee, C.; Yang, W.; Parr, R. G. Phys. ReV. B: Condens. Matter Mater. Phys. 1988, 37, 785.
(36) Toulouse, J.; Savin, A.; Adamo, C. J. Chem. Phys. 2002, 117, 10465.
(4) Perdew, J. P. In Electronic Structure of Solids ‘91; Ziesche, P., Eschig, H., Eds.; Akademie Verlag: Berlin, 1991; p 11.
(37) Boese, A. D.; Martin, J. M. L.; Handy, N. C. J. Chem. Phys. 2003, 119, 3005.
(5) Perdew, J. P.; Wang, Y. Phys. ReV. B: Condens. Matter Mater. Phys. 1992, 45, 13244.
(38) Becke, A. D. J. Chem. Phys. 2003, 119, 2972.
(6) Becke, A. D. J. Chem. Phys. 1993, 98, 1372.
(39) Boese, A. D.; Chandra, A.; Martin, J. M. L.; Marx, D. J. Chem. Phys. 2003, 119, 5965.
(7) Becke, A. D. J. Chem. Phys. 1993, 98, 5648. (8) Stephens, P. J.; Devlin, F. J.; Chabalowski, C. F.; Frisch, M. J. J. Phys. Chem. 1994, 98, 11623. (9) Gill, P. M. W. Mol. Phys. 1996, 89, 433.
(40) Guner, V.; Khuong, K. S.; Leach, A. G.; Lee, P. S.; Bartberger, M. D.; Houk, K. N. J. Phys. Chem. A 2003, 107, 11445.
(10) Becke, A. D. J. Chem. Phys. 1996, 104, 1040.
(41) Tao, J.; Perdew, J. P.; Staroverov, V. N.; Scuseria, G. E. Phys. ReV. Lett. 2003, 91, 146401.
(11) Perdew, J. P.; Burke, K.; Ernzerhof, M. Phys. ReV. Lett. 1996, 77, 3865.
(42) Staroverov, V. N.; Scuseria, G. E.; Tao, J.; Perdew, J. P. J. Chem. Phys. 2003, 119, 12129.
(12) Perdew, J. P.; Ernzerhof, M.; Burke, K. J. Chem. Phys. 1996, 105, 9982.
(43) Li, Q. S.; Xu, X. D.; Zhang, S. Chem. Phys. Lett. 2004, 384, 20.
(13) Becke, A. D. J. Chem. Phys. 1997, 107, 8554.
(44) Staroverov, V. N.; Scuseria, G. E.; Tao, J.; Perdew, J. P. Phys. ReV. B: Condens. Matter Mater. Phys. 2004, 69, 75102.
(14) Adamo, C.; Barone, V. Chem. Phys. Lett. 1997, 274, 242. (15) Adamo, C.; Barone, V. J. Chem. Phys. 1998, 108, 664. (16) Schmider, H. L.; Becke, A. D. J. Chem. Phys. 1998, 108, 9624. (17) Rey, J.; Savin, A. Int. J. Quantum Chem. 1998, 69, 581.
(45) Perdew, J. P.; Tao, J.; Staroverov, V. N.; Scuseria, G. E. J. Chem. Phys. 2004, 120, 6898. (46) Zhao, Y.; Pu, J.; Lynch, B. J.; Truhlar, D. G. Phys. Chem. Chem. Phys. 2004, 6, 673.
Design of Density Functionals (47) Xu, X.; Goddard, W. A. Proc. Natl. Acad. Sci. U.S.A. 2004, 101, 2673.
J. Chem. Theory Comput., Vol. 2, No. 2, 2006 381
(48) Coote, M. L. J. Phys. Chem. 2004, 108, 3865.
(75) Champagne, B.; Perpete, E. A.; Jacquemin, D.; Gisbergen, S. J. A. V.; Baerends, E.-J.; Soubra-Ghaoui, C.; Robins, K. A.; Kirtman, B. J. Phys. Chem. A 2000, 104, 4755.
(49) Zhao, Y.; Lynch, B. J.; Truhlar, D. G. J. Phys. Chem. A 2004, 108, 2715.
(76) Bulat, F. A.; Toro-Labbe´, A.; Champagne, B.; Kirtman, B.; Yang, W. J. Chem. Phys. 2005, 123, 14319.
(50) Boese, A. D.; Martin, J. M. L. J. Chem. Phys. 2004, 121, 3405.
(77) Langhoff, S. R.; Bauschlicher, C. W. Annu. ReV. Phys. Chem. 1988, 39, 181.
(51) Zhao, Y.; Truhlar, D. G. J. Phys. Chem. A 2004, 108, 6908.
(78) Widmark, O.; Malmqvist, P. A.; Roos, B. Theor. Chim. Acta 1990, 77, 291.
(52) Andersson, S.; Gruning, M. J. Phys. Chem. A 2004, 108, 7621. (53) Johnson, E. R.; Wolkow, R. A.; DiLabio, G. A. Chem. Phys. Lett. 2004, 394, 334. (54) Zhao, Y.; Lynch, B. J.; Truhlar, D. G. Phys. Chem. Chem. Phys. 2005, 7, 43. (55) Zhao, Y.; Gonza´lez-Garcı´a, N.; Truhlar, D. G. J. Phys. Chem. A 2005, 109, 2012. (56) Zhao, Y.; Truhlar, D. G. J. Chem. Theory Comput. 2005, 1, 415. (57) Schultz, N.; Zhao, Y.; Truhlar, D. G. J. Phys. Chem. A 2005, 109, 4388. (58) Zhao, Y.; Truhlar, D. G. J. Phys. Chem. A 2005, 109, 5656. (59) Zhao, Y.; Truhlar, D. G. J. Phys. Chem. A 2005, 109, 6624. (60) Zhao, Y.; Truhlar, D. G. Phys. Chem. Chem. Phys. 2005, 7, 2701.
(79) Pou-Amerigo, R.; Merchan, M.; Nebot-Gil, I.; Widmark, P. O.; Roos, B. Theor. Chim. Acta 1995, 92, 149. (80) Pople, J. A.; Head-Gordon, M.; Raghavachari, K. J. Chem. Phys. 1987, 87, 5968. (81) Fast, P. L.; Sanchez, M. L.; Truhlar, D. G. Chem. Phys. Lett. 1999, 306, 407. (82) Curtiss, L. A.; Redfern, P. C.; Raghavachari, K.; Rassolov, V.; Pople, J. A. J. Chem. Phys. 1999, 110, 4703. (83) Curtiss, L. A.; Raghavachari, K.; Redfern, P. C.; Rassolov, V.; Pople, J. A. J. Chem. Phys. 1998, 109, 7764. (84) Frisch, M. J.; Pople, J. A.; Binkley, J. S. J. Chem. Phys. 1984, 80, 3265. (85) Hehre, W. J.; Radom, L.; Schleyer, P. v. R.; Pople, J. A. Ab Initio Molecular Orbital Theory; Wiley: New York, 1986. (86) Fast, P. L.; Truhlar, D. G. J. Phys. Chem. A 2000, 104, 6111.
(61) Perdew, J. P.; Ruzsinszky, A.; Tao, J.; Staroverov, V. N.; Scuseria, G. E.; Csonka, G. I. J. Chem. Phys. 2005, 123, 62201.
(87) Sinnokrot, M. O.; Sherrill, C. D. J. Phys. Chem. A 2004, 108, 10200.
(62) Becke, A. D. J. Chem. Phys. 2005, 122, 64101.
(88) Stevens, W.; Basch, H.; Krauss, J. J. Chem. Phys. 1984, 81, 6026.
(63) Dickson, R. M.; Becke, A. D. J. Chem. Phys. 2005, 123, 111101. (64) Keal, T. W.; Tozer, D. J. J. Chem. Phys. 2005, 123, 121103. (65) Zhao, Y.; Schultz, N. E.; Truhlar, D. G. J. Chem. Phys. 2005, 123, 161103. Note that, in this communication, we interchanged cCRβ,i and cCσσ,i in Table 1. In addition, “reduced density xσ” before eq 1 should read "reduced density gradient xσ”. (66) Schultz, N.; Zhao, Y.; Truhlar, D. G. J. Phys. Chem. A 2005, 109, 11127. (67) Quintal, M. M.; Karton, A.; Iron, M. A.; Boese, A. D.; Martin, J. M. L. J. Phys. Chem. A [ASAP article] 2006, 110 (2), 709-716. (68) Lynch, B. J.; Truhlar, D. G. J. Phys. Chem. A 2003, 107, 8996. (69) Chakravorty, S. J.; Gwaltney, S. R.; Davidson, E. R.; Parpia, F. A.; Fischer, C. F. F. Phys. ReV. A: At., Mol., Opt. Phys. 1993, 47, 3649. (70) Lynch, B. J.; Zhao, Y.; Truhlar, D. G. J. Phys. Chem. A 2003, 107, 1384. (71) Izgorodina, E. I.; Coote, M. L.; Radom, L. J. Phys. Chem. A 2005, 109, 7558. (72) Lynch, B. J.; Truhlar, D. G. J. Phys. Chem. A 2003, 107, 3898. (73) Lynch, B. J.; Zhao, Y.; Truhlar, D. G. J. Phys. Chem. A 2005, 109, 1643. (74) Scott, A. P.; Radom, L. J. Phys. Chem. 1996, 100, 16502.
(89) Boys, S. F.; Bernardi, F. Mol. Phys. 1970, 19, 553. (90) Schwenke, D. W.; Truhlar, D. G. J. Chem. Phys. 1985, 82, 2418. (91) Frisch, M. J.; Trucks, G. W.; Schlegel, H. B.; Scuseria, G. E.; Robb, M. A.; Cheeseman, J. R.; Montgomery, J. A., Jr.; Kudin, K. N.; Burant, J. C.; Millam, J. M.; Iyengar, S. S.; Tomasi, J.; Barone, V.; Mennucci, B.; Cossi, M.; Scalmani, G.; Rega, N.; Petersson, G. A.; Nakatsuji, H.; Hada, M.; Ehara, M.; Toyota, K.; Fukuda, R.; Hasegawa, J.; Ishida, M.; Nakajima, T.; Honda, Y.; Kitao, O.; Nakai, H.; Klene, M.; Li, X.; Knox, J. E.; Hratchian, H. P.; Cross, J. B.; Adamo, C.; Jaramillo, J.; Gomperts, R.; Stratmann, R. E.; Yazyev, O.; Austin, A. J.; Cammi, R.; Pomelli, C.; Ochterski, J. W.; Ayala, P. Y.; Morokuma, K.; Voth, G. A.; Salvador, P.; Dannenberg, J. J.; Zakrzewski, G.; Dapprich, S.; Daniels, A. D.; Strain, M. C.; Farkas, O.; Malick, D. K.; Rabuck, A. D.; Raghavachari, K.; Foresman, J. B.; Ortiz, J. V.; Cui, Q.; Baboul, A. G.; Clifford, S.; Cioslowski, J.; Stefanov, B. B.; Liu, G.; Liashenko, A.; Piskorz, P.; Komaromi, I.; Martin, R. L.; Fox, D. J.; Keith, T.; Al-Laham, M. A.; Peng, C. Y.; Nanayakkara, A.; Challacombe, M.; Gill, P. M. W.; Johnson, B.; Chen, W.; Wong, M. W.; Gonzalez, C.; Pople, J. A. Gaussian 03, revision C.01; Gaussian, Inc.: Pittsburgh, PA, 2003. (92) Werner, H.-J.; Knowles, P. J.; Amos, R. D.; Bernhardsson, A.; Berning, A.; Celani, P.; Cooper, D. L.; Deegan, M. J. O.; Dobbyn, A. J.; Eckert, F.; Hampel, C.; Hetzer, G.; Korona, T.; Lindh, R.; Lloyd, A. W.; McNicholas, S. J.;
382 J. Chem. Theory Comput., Vol. 2, No. 2, 2006 Manby, F. R.; Meyer, W.; Mura, M. E.; Nicklass, A.; Palmieri, P.; Pitzer, R.; Rauhut, G.; Schu¨tz, M.; Schumann, U.; Stoll, H.; Stone, A. J.; Tarroni, R.; Thorsteinsson, T. MOLPRO, 2002.6; University of Birmingham: Birmingham, AL, 2002. (93) Stoll, H.; Pavkidou, C. M. E.; Preuss, H. Theor. Chim. Acta 1978, 49, 143.
Zhao et al. (118) Fiacco, D. L.; Mo, Y.; Hunt, S. W.; Ott, M. E.; Roberts, A.; Leopold, K. R. J. Phys. Chem. A 2001, 105, 484. (119) Hartree, D. R.; Hartree, W.; Swirles, B. Philos. Trans. R. Soc. London, Ser. A 1939, 238, 229. (120) Anderson, W. B.; Burdett, J. K.; Czech, P. T. J. Am. Chem. Soc. 1994, 116, 8808. (121) McKoy, V.; Sinanoglu, V. J. Chem. Phys. 1964, 41, 2689.
(94) Gori-Giorgi, P.; Sacchetti, F.; Bachelet, G. G. Phys. ReV. B: Condens. Matter Mater. Phys. 2000, 61, 7353.
(122) Laidig, W. D.; Schaefer, H. F. J. Chem. Phys. 1981, 74, 3411.
(95) Gori-Giorgi, P.; Perdew, J. P. Phys. ReV. B: Condens. Matter Mater. Phys. 2004, 69, 041103.
(123) Lee, T. J.; Taylor, P. R. Int. J. Quantum Chem., Quantum Chem. Symp. 1989, 23, 199.
(96) von-Weizsa¨cker, C. F. Z. Phys. 1935, 96, 431.
(124) Das, G.; Wahl, A. C. J. Chem. Phys. 1967, 47, 2934.
(97) Yang, G.; Reinstein, L. E.; Pai, S.; Xu, Z.; Carroll, D. L. Med. Phys. 1998, 25, 2308.
(125) Botch, B. H.; Dunning, T. H.; Harrison, J. F. J. Chem. Phys. 1981, 47, 2934.
(98) Levy, M.; Perdew, J. P. Phys. ReV. A: At., Mol., Opt. Phys. 1985, 32, 2010.
(126) Walch, S. P.; Bauschlicher, C. W. Chem. Phys. Lett. 1982, 86, 66.
(99) Lieb, E. H.; Oxford, S. Int. J. Quantum Chem. 1981, 19, 427.
(127) Chong, D. P.; Langhoff, S. R. J. Chem. Phys. 1986, 84, 5606.
(100) Dybala-Defratyka, A.; Paneth, P.; Pu, J.; Truhlar, D. G. J. Phys. Chem. A 2004, 108, 2475. (101) Curtiss, L. A.; Raghavachari, K.; Redfern, P. C.; Pople, J. A. J. Chem. Phys. 2000, 112, 7374. (102) Henry, D. J.; Sullivan, M. B.; Radom, L. J. Phys. Chem. A 2003, 118, 4849. (103) Hautman, J.; Klein, M. L. NATO ASI Ser., Ser. E 1991, 205, 395. (104) Karlin, K. D. Science 1993, 261, 701. (105) Crabtree, R. H. The Organometallic Chemistry of the Transition Metals, 2nd ed.; Wiley: New York, 1994. (106) George, S. M. Chem. ReV. 1995, 95, 475. (107) Somorjai, G. A. Chem. ReV. 1995, 96, 1223. (108) Ratner, M. A.; Davis, B.; Kemp, M.; Mujica, V.; Roitberg, A.; YalirakiI, S. Ann. N. Y. Acad. Sci. 1998, 852, 22. (109) Transition State Modeling for Catalysis; Truhlar, D. G., Morokuma, K., Eds.; ACS Symposium Series 721; American Chemical Society: Washington, DC, 1999.
(128) Carter, E. A.; Goddard, W. A. J. Chem. Phys. 1988, 88, 1752. (129) Murthy, R. B.; Messemer, R. P. J. Chem. Phys. 1992, 97, 4974. (130) Schmidt, M. W.; Gordon, M. S. Annu. ReV. Phys. Chem. 1998, 49, 233. (131) Staroverov, V. N.; Davidson, E. R. J. Am. Chem. Soc. 2000, 122, 186. (132) Gra¨fenstein, J.; Cremer, D. Chem. Phys. Lett. 2000, 316, 569. (133) Pollet, R.; Savin, A.; Leininger, T.; Stoll, H. J. Chem. Phys. 2002, 116, 1250. (134) Dunning, T. H.; Harding, L. B. In Theory of Chemical Reaction Dynamics; Baer, M., Ed.; CRC press: Boca Raton, FL, 1985; p 1. (135) Brown, F. B.; Steckler, R.; Schwenke, D. W.; Truhlar, D. G.; Garrett, B. C. J. Chem. Phys. 1985, 82, 188. (136) Perdew, J. P.; Zunger, A. Phys. ReV. B: Condens. Matter Mater. Phys. 1981, 23, 5048.
(110) Davidson, E. R. Chem. ReV. 2000, 100, 351.
(137) Kryachko, E. S.; Luden˜a, E. V. Energy Density Functional Theory of Many-Electron Systems; Kluwer Academic: Dordrecht, The Netherlands, 1990; p 637.
(111) Siegbahn, P. E. M.; Blomberg, M. R. A. Chem. ReV. 2000, 100, 421.
(138) Ciofini, I.; Adamo, C.; Chermette, H. J. Chem. Phys. 2005, 123, 121102.
(112) Gladysz, J. A. Chem. ReV. 2000, 100, 1167.
(139) Dreuw, A.; Head-Gordon, M. Chem. ReV. 2005, 105, 4009.
(113) Handbook on Metalloproteins; Bertini, I.; Sigel, A.; Sigel, H., Eds.; Dekker: New York, 2001.
(140) Slater, J. C. Phys. ReV. 1954, 528.
(114) Rappe, A. K.; Skiff, W. M.; Casewit, C. J. Chem. ReV. 2000, 100, 1435.
(142) Ziegler, T. Chem. ReV. 1991, 91, 651.
(115) Coperat, C.; Chabonas, M.; Saint-Arromon, R. P.; Basset, J.-M. Angew. Chem., Int. Ed. 2003, 42, 156.
(143) Gritsenko, O. V.; Schipper, P. R. T.; Baerends, E. J. J. Chem. Phys. 1997, 107, 5007.
(116) Phillips, J. A.; Cramer, C. J. J. Chem. Theory Comput. 2005, 1, 827.
(144) Slater, J. C. Quantum Theory of Molecules and Solids. Vol. 4: The Self-Consistent Field for Molecules and Solids; McGraw-Hill: New York, 1974.
(117) Reeve, S. W.; Burns, W. A.; Lovas, F. J.; Suenram, R. D.; Leopold, K. R. J. Phys. Chem. A 1993, 97, 10630.
(141) Tschinke, V.; Ziegler, T. A. J. Chem. Phys. 1990, 93, 8051.
CT0502763